API Overview / Synopses

MANUAL NAVIGATION: Preceding Page - Next Page - Table of Contents - Reference

---

Here we show how Flow-IPC fulfills its aims; let's take a high-level tour through the available API. (The preceding page (Manual Start) Flow-IPC: Bird's-eye View summarizes the aims of Flow-IPC.)

We provide a synopsis in ~each section below, to give a sense of what your code using that feature would look like. It's both a survey and a cheat sheet, but it is not comprehensive the way the individual subsequent pages, or the Reference, are. Please jump to a particular page to drill down on any topic of interest (Table of Contents).

Sessions

No matter the particular IPC transport (Unix domain socket, MQ, SHM pool), an ever-present need is to establish a naming scheme and coordinate it between the participant processes. To connect one Unix domain socket, one needs to name a server socket, and the connecting process needs to know that name. An MQ requires a name, as does a SHM pool. Developing and maintaining a convention is, normally, both necessary and annoying. Instead:

There is a simple scheme that designates one side of a conversation as the session server, the other a client; the latter connects to the former. (A session server process can carry on multiple sessions with different client processes.) The user need only specify the basic facts of the IPC-participating applications (like executable names and owners, and which ones are servers, and which clients they'll accept as the opposing side). This is done via the straightforward struct ipc::session::App + its children. The rest of ipc::session establishes internal naming conventions based off that information. Having specified those App basics, the user opens (on each side) an equally capable ipc::session::Session that is the context for all subsequent IPC between those 2 processes.

Having obtained a Session, the application can open transport channels (and, if desired, SHM arenas) without any naming decisions to make. In fact, in most situations, one can open the session and any desired channels at the same time (i.e., a certain number of channels are ready to use on each side, once the Session is opened).

Synopsis: Sessions

‍Describe the universe of communicating apps via simple structs:

using namespace ipc::session;
 
// Example: 2 applications.
const App APP_A{ "a", // Unique name for identification in the system.
                 // Basic properties for safety, authentication.
                 "/bin/app_a.exec", USER_ID_A, GROUP_ID };
const App APP_B{ "b", "/bin/app_b.exec", USER_ID_B, GROUP_ID };
// Which app(s) will connect sessions (client) versus accept sessions (server)?
const Client_app APP_B_AS_CLI{ APP_B };
const Server_app APP_A_AS_SRV{ APP_A,
                               { APP_B.m_name }, // We are A; we allow only B to open sessions with us.
                               "", // Keep run-time files in /var/run by default.
                               // Safety/auth preset.  No need for `0644` and so forth.
                               ipc::util::Permissions_level::S_GROUP_ACCESS };
// Collect the client apps.  (These objects must outlive any Flow-IPC objects to which they are passed.)
const Client_app::Master_set MASTER_APPS_AS_CLI{ { APP_B_AS_CLI.m_name, APP_B_AS_CLI } };

Compile-time-configure your sessions:

// Specify zero-copy and SHared Memory (SHM) abilities if any.
 
namespace ssn = shm::classic;
// Use [ipc::]session - For non-SHM-enabled communication.
//     [ipc::session::]shm::* - For SHM-enabled comms (zero-copy structured messaging; explicit SHM access if desired).
//     [ipc::session::]shm::classic - Classic one-segment variant with simple allocation strategy by Boost.
//                     shm::arena_lend::jemalloc - Adaptive multi-segment variant with allocation by jemalloc.
// Subsequent API remains ~the same.  Only your convenience alias above changes.
 
// Specify the low-level transports used underneath in your future IPC channels.
 
// Example: Your future channels will use Unix domain sockets (a good default; see the channel-opening chapter
//   for alternatives), and they can transmit FDs (network socket, file handle, etc.).
constexpr auto MQ_TYPE = schema::MqType::NONE;
constexpr bool TRANSMIT_FDS = true;
 
// Client side:
using Session = ssn::Client_session<MQ_TYPE, TRANSMIT_FDS>;
// Server side:
using Session_server = ssn::Session_server<MQ_TYPE, TRANSMIT_FDS>;
using Session = Session_server::Server_session_obj;

Open session(s):

// Client side:
Session session{nullptr, APP_B_AS_CLI, // This is us (connector).
                APP_A_AS_SRV, // The opposing application (acceptor).
                ...};
Error_code err_code;
session.sync_connect(&err_code); // Synchronous, non-blocking.
if (!err_code) { ...`session` is ready.... } else { ...opposing server not active/listening... }
 
// Server side:
Session_server session_srv{nullptr, APP_A_AS_SRV, // This is us (acceptor).
                           MASTER_APPS_AS_CLI}; // These are potential opposing apps.
Session session;
session_srv.async_accept(&session,
                         [...](const auto& err_code) { if (!err_code) { ...`session` is ~ready.... } });
// ^-- asio-style API.  Non-blocking/synchronous API also available (integrate directly with poll(), epoll_...(), etc.).
 
// NOTE: Upon opening session, capabilities of `session` on either side are **exactly the same**.
// Client/server status matters only when establishing the IPC conversation;
// the conversation itself once established is arbitrary and up to you fully.

Open channel(s) in a session:

• Easiest: Have them be pre-opened, so that they're ready from the session start:

// Client side: Example: Expect 1 client-requested channel; N server-requested channels.
Session::Channels my_init_channels{1};
Session::Channels their_init_channels;
Error_code err_code;
session.sync_connect(..., &my_init_channels, ..., &their_init_channels, &err_code); // Synchronous, non-blocking.
if (!err_code)
{
  // `session` is ready, but we might not even care that much because:
  //   my_init_channels[0] = The 1 pre-opened Session::Channel_obj we requested.
  //   their_init_channels = vector<Session::Channel_obj> = The N pre-opened channels opposing side requested.
  // Sessions are mostly a way to establish channels after all, and we've got ours!
}
// Server side: Example: Expect <cmd line arg> server-requested channels; N client-requested channels.
Session::Channels my_init_channels{lexical_cast<size_t>(argv[2])};
Session::Channels their_init_channels;
session_srv.async_accept(&session, ..., &my_init_channels, ..., &their_init_channels, ..., ...,
                         [...](const auto& err_code)
{
  if (!err_code)
  {
    // session = Opened session.
    // my_init_channels = The pre-opened `Session::Channel_obj`s we requested (argv[2] of them).
    // their_init_channels = The N pre-opened channels opposing side requested.  (N=1 in this example; see above.)
  }
});

• Max flexibility: Open them at any time. Either side can be initiator or acceptor of channel opening at any time.

// session.open_channel() is non-blocking, synchronous.
 
// Active opener (side 1):
Session::Channel_obj channel;
session.open_channel(&channel); // Non-blocking, synchronous.
// Active opener advanced form (allows to pass 1 structured message along with the channel-open action):
auto mdt = session.mdt_builder();
mdt->setCoolParameterFromMyCapnpSchema(42.42);
session.open_channel(&channel, std::move(mdt));
 
// Passive opener (side 2):
session.init_handlers(..., [...](Session::Channel_obj&& channel, auto&& mdt)
{
  // `channel` is ready.  Also the structured message (if desired) is immediately available:
  const float val = mdt->getCoolParameterFromMyCapnpSchema();
  std::cout << "Along with the channel-open request, other guy sent us this number: [" << val << "].\n";
});

Once you've got your session and/or channels, it's off to the races... namely:

Transport (unstructured)

A channel is a bundling of the resources required for, essentially, a bidirectional pipe capable of transmitting binary messages (blobs) and, optionally, native handles. A particular ipc::transport::Channel (as specified at compile-time via its template parameters) may, for example, consist of an outgoing POSIX MQ (message queue) handle and a similar incoming-MQ handle; or conceptually similar SHM-backed MQs (boost.ipc MQs). Alternatively it can consist of a Unix domain socket endpoint (which is bidirectional). (Other combinations are generically possible.)

If one uses the ipc::session API, one need not worry about any detail of opening these various handles and feeding them to the particular Channel. This is shown in the Sessions synopsis above. The type of channel openable via a given session is compile-time-configured when declaring your convenience Session alias. The Sessions synopsis above makes one example selection. You can make other choices as follows. Note that you just make these selections – the details are handled underneath. E.g., you need no knowledge of POSIX MQ API (from man mq_overview) to take advantage of its performance characteristics.

Synopsis: Pick channel low-level transport

‍There are two knobs to twiddle. If unsure about the first knob, NONE is fine.

constexpr auto MQ_TYPE = schema::MqType::NONE; // Do not use MQs; always use Unix domain sockets.  (A good default.)
                       = schema::MqType::POSIX; // Use POSIX MQs (as from `man mq_overview`) when possible.
                       = schema::MqType::BIPC; // Use `boost::interprocess::message_queue`s when possible.
constexpr bool TRANSMIT_FDS = true; // You need ability to send native handles (FDs).  Use Unix domain sockets when needed.
                            = false; // You don't need this ability.  So Unix domain socket used only if MQ_TYPE=NONE.
// Client:
using Session = ssn::Client_session<MQ_TYPE, TRANSMIT_FDS>;
// Server:
using Session_server = ssn::Session_server<MQ_TYPE, TRANSMIT_FDS>;
using Session = Session_server::Server_session_obj;

Alternatively a simple application or prototype may want to manually set them up without ipc::session. Though, in our humble opinions, using ipc::session is what's simple, as manual setup of a channel involves major annoyances including naming and transport-specific considerations (Unix domain sockets versus POSIX MQs versus... etc.). Perhaps the better term would have been "conventional" rather than "simple." We will briefly discuss how to manually open various channels without ipc::session a few paragraphs below.

Note

Note well! Once you've obtained an open transport::Channel, whether via ipc::session or manually, in many cases you will not need or want to understand the information in the rest of this section and instead skip straight to Transport (structured), below. That is, often you'll want to transmit structured data – various fields, structures, lists thereof, etc. (+ possibly FDs) – as opposed to mere binary blobs (+ possibly FDs). To do so, one takes a freshly-opened Channel and upgrades it to a structured channel (struc::Channel). A good portion of the power of Flow-IPC is how it integrates with/leverages Cap'n Proto via ipc::transport::struc::Channel. However, for closer-to-the-metal use cases it is completely reasonable to want to operate at the lower, unstructured layer that is ipc::transport::Channel, or even the components it comprises (MQs, stream sockets at least). In that case, and/or to see how to manually open these, read on in this section.

All transport APIs at this layer, as well as the structured layer (see below), have an API that is as-synchronous-as-possible. In particular all send operations are non-blocking and synchronous and never return "would-block" – while guaranteeing good performance in practice. Receive APIs are asynchronous by their nature. A proactor-style API (a-la boost.asio) is provided for this and other async ops (such as connect and accept).

Note

An alternative API is available for all objects with any async operations (generally: any subset of send, receive, connect, accept). This is called the sync_io pattern, contrasting with the above async-I/O pattern. The sync_io-pattern alternative for each given I/O object is similar in spirit to reactor-style (non-blocking) APIs a-la OpenSSL's. It may help you integrate with a reactor-style event loop such as one built on epoll() or poll(). It is also the way to strictly control the background threads in your application arising for Flow-IPC's operation: if desired you can have it operate entirely within your own threads and event loop(s). See Asynchronicity and Integrating with Your Event Loop for more discussion of this somewhat-hairy topic.

This async-I/O-versus-sync_io choice is available for all other parts of Flow-IPC, including: ipc::transport structured layer (struc::sync_io::Channel versus just struc::Channel for async-I/O alternative) and ipc::session.

Synopsis: Transport (unstructured)

‍Lowest layer: MQs accessible via unified ipc::transport::Persistent_mq_handle concept. (These APIs are mutually uniform, more pleasant, and at times more powerful than the underlying APIs.) boost.asio is leveraged for Unix domain sockets. We use these internally, but we expose them publicly too, should you choose to work at that low level.

• Posix_mq_handle for POSIX message queues.
• Bipc_mq_handle for boost.interprocess MQs – internally SHM-based.
• We leverage boost::asio::local::seq_packet_protocol::socket, aliased from ipc::transport::asio_local_stream_socket::Peer_socket. Namespace asio_local_stream_socket has advanced utilities.
  • Some internal-impl information of interest: In Linux, at least, this is known as the SOCK_SEQPACKET variety of local-domain (Unix-domain) socket, which is connection-oriented in the same way as the far-more-ubiquitous SOCK_STREAM variety but has datagram-oriented semantics like SOCK_DGRAM. This allows for significant perf gains (in both the structured and unstructured layers) due to message boundaries being handled by the kernel; and further gains by making batch-receiving natively available to us. It is possible to switch to SOCK_STREAM by flipping a compile-time constant in the source code.
  • Class template asio_local_stream_socket::Msg_batch_in provides stateful high-performance batch-reading (internally, "::recvmmsg()") over a local socket.

Next layer: Sender/receiver concepts and their implementations. The key concepts:

• ipc::transport::Blob_sender:
  • Method send_blob() to synchronously, never-would-blockingly send a binary buffer with message boundaries respected.
  • Methods *end_sending() to send an end-of-stream graceful-close marker.
  • Method auto_ping() to enable periodic auto-pinging of opposing receiver with internal messages that are ignored except that they reset any idle timer as enabled via opposing Blob_receiver::idle_timer_run().
• ipc::transport::Blob_receiver:
  • Method async_receive_blob() to receive, with message boundaries respected, one message the opposing side sent via send_blob(), into a large-enough buffer supplied by caller. Receipt, or error, is indicated asynchronously via completion handler function supplied by user.
    • Method async_receive_blob_batch() to batch-receive up-to-N such messages via a reusable ipc::transport::Msg_batch_in object. (This may result in significant perf gains, when the concept impl is able to leverage a native batch-receiving capability. This is the case with ipc::transport::Native_socket_stream. There shall be ~no perf loss, even when there is no native batch-receiving capability available, as is the case with ipc::transport::Blob_stream_mq_receiver.)
  • Method idle_timer_run() to enable an idle timer that would emit a connection-ending error, if no message (whether from a user blob or an auto_ping()) arrives within a specific time period from the last such message or connection start.
• ipc::transport::Native_handle_sender: Just like Blob_sender, except send_blob() is replaced by:
  • Method send_native_handle() which sends: a message (blob); or a socket handle (FD); or both. User provides socket handle (if any) as effectively an int.
• ipc::transport::Native_handle_receiver: What Blob_receiver is to Blob_sender, this guy is to Native_handle_sender. Method async_receive_native_handle() can receive a message, an FD, or both; and accordingly async_receive_native_handle_batch() can batch-receive same.
• For each ipc::transport::X concept above, ipc::transport::sync_io::X is the sync_io-pattern counterpart. E.g.: ipc::transport::sync_io::Blob_receiver allows for integration of blob-receiving functionality above into an epoll-based event loop, with only non-blocking calls and no internal background threads involved.

These classes or class templates implement the above concepts:

• Unix domain sockets have exclusive FD-transmission ability; are great for blobs too.
  • ipc::transport::Native_socket_stream implements all 4 concepts: Blob_sender, Blob_receiver, Native_handle_sender, Native_handle_receiver.
• MQs efficiently transmit blobs of limited size. One kernel-persistent MQ object, maintained by OS, is required for a unidirectional pipe. For bidirectional pipe, use 2 MQs arranged in mutually-facing fashion. For each one, use a sender object on one end and receiver on another.
  • ipc::transport::Blob_stream_mq_sender implements Blob_sender.
  • ipc::transport::Blob_stream_mq_receiver implements Blob_receiver.

The concept APIs above act the same regardless of concrete class and apply to each object once in opened (PEER) state. Each class also has non-concept aspects which depend on its specific characteristics. Most of these concern opening a connection (getting to PEER state). Also ipc::transport::Native_socket_stream can leverage the OS's ability to obtain credentials (user ID, group ID, process ID) of the opposing process.

A survey:

using namespace ipc::transport;
namespace util = ipc::util;
namespace asio_local = ipc::transport::asio_local_stream_socket::local_ns;
 
// With MQs, you must open an MQ handle first, which can create (if necessary) the kernel-persistent OS MQ object.
 
// ipc::session would figure all this out by itself; when working manually you have to do so yourself.
Posix_mq_handle out_q{nullptr, util::Shared_name::ct("sharedMqNameA2B"), // Shared name known by other side
                      util::OPEN_OR_CREATE, // Atomically create if needed, open if exists in system already.
                      10, 1000, // Max # of unread messages, max size of message,
                      0600, &err_code}; // Permissions, if creation is needed; Error_code* in case of error.
Bipc_mq_handle in_q{nullptr, util::Shared_name::ct("sharedMqNameB2A"), util::OPEN_ONLY, &err_code};
 
// Then to make yourself a Blob_sender/receiver, feed them into constructor.
Blob_stream_mq_sender<Posix_mq_handle> out{nullptr, "cool_reader_nickname", std::move(out_q)};
Blob_stream_mq_receiver<Bipc_mq_handle> in{nullptr, "cool_writer_nickname", std::move(in_q)};
 
// With socket streams, you can either use two pre-opened FDs; or use client/server setup.
 
// Pre-opened FDs:
asio_local::Peer_socket a_raw_sock, b_raw_sock;
asio_local::connect_pair(a_raw_sock, b_raw_sock); // Can also use native ::socketpair().
// (b_raw_sock could be spread to another process via fork() or another Native_socket_stream and other techniques.)
Native_socket_stream a_sock{nullptr, "cool_a_nickname", util::Native_handle{a_raw_sock}};
Native_socket_stream b_sock{nullptr, "cool_b_nickname", util::Native_handle{b_raw_sock}};
 
// Client/server: client.
const auto SRV_ADDR = util::Shared_name::ct("sharedSockNameSrv");
// ^-- Client and server need to agree on this, and it must not conflict.  ipc::session would take care of this.
Native_socket_stream sock{nullptr, "cool_a_nickname"};
Error_code err_code;
sock.sync_connect(SRV_ADDR, &err_code); // Synchronous, non-blocking.
 
// Client/server: server.  A dedicated server-acceptor class exists for this.
 
Native_socket_stream sock{nullptr, "cool_b_nickname"};
Native_socket_stream_acceptor srv{nullptr, SRV_ADDR, &err_code}; // Error_code* to catch, e.g., name conflicts.
srv.async_accept(&sock, [...](const auto& err_code) { ... });
 
// Native_socket_stream adds some safety/informational features.  ipc::session uses this internally.
 
const util::Process_credentials& creds = sock.remote_peer_process_credentials();
std::cout << creds // Prints .process_id(), .user_id(), .group_id()
          << ' ' << creds.process_invoked_as() << '\n'; // Prints the command-line argv[0] through which remote process was invoked.
 
// Can bundle Blob_senders, Blob_receivers, etc., into `ipc::transport::Channel`s.
// Convenience aliases and data-free subclasses exist to make it easier versus using an actual Channel<...>.
 
// In this case we create a Channel with 2 bidirectional pipes, from 2 `Posix_mq_handle`s like out_q above; and
// an already-opened Native_socket_stream.
Mqs_socket_stream_channel<false, Posix_mq_handle> // false => async-I/O API; true => `sync_io`-pattern API.
  chan{nullptr, "two_pipe_chan_nickname", std::move(out_q), std::move(in_q), std::move(sock)};
// Use the blobs pipe to send:
chan.send_blob(util::Blob_const{some_byte_vector.data(), some_byte_vector.size()});
// Use the blobs-and-FDs pipe to receive:
util::Native_handle fd;
chan.async_receive_native_handle(&fd, util::Blob_mutable{some_target_vector.data(), some_target_vector.capacity()},
                                 [&](const auto& err_code, size_t n_rcvd)
{
  if (!err_code)
  {
    some_target_vector.resize(n_rcvd);
    // some_target_vector.begin()..end() = received blob.
    // fd.m_native_handle = `int` file descriptor/raw handle.
    // ...Deal with those inputs....
  }
});
// See also `chan.async_receive_native_handle_batch()` for (optional) batch-receiving: multiple in-messages per
// system call, where natively supported (as with Native_socket_stream).  struc::Channel uses it automatically when
// the transport recommends it.  To use it directly see the Native_handle_receiver concept doc, "Batch-receiving."
// (To do this natively, in Linux with Unix domain sockets, one must wrangle `::recvmmsg()` which in itself is not
// so easy.  Furthermore, to get the presumably desired perf benefit, it is ~necessary to maintain a message array
// across multiple such calls, recycling untouched entries and replacing the others, while avoiding linear-time
// ops: not easy at all, with many potential pitfalls.
// Native_socket_stream::async_receive_{blob|native_handle}_batch() and
// class Native_socket_stream::{Blob|Native_handle}_batch_in (impl of Msg_batch_in) are the answer to these
// challenges.)
 
// In this case we create a Channel with 1 bidirectional pipe, a Unix domain socket stream, for both FDs and blobs.
Socket_stream_channel<false> chan{..., std::move(sock)};
// Can access the Native_handle_sender through the accessor:
chan.hndl_snd()->send_native_handle(util::Native_handle{a_raw_int_fd}, util::Blob_const{...});
// Or through the Channel forwarding API:
chan.send_native_handle({}, util::Blob_const{...}); // Send a blob but no FD this time.
 
// Or create some `Channel`s with a blobs-only pipe each:
Socket_stream_channel_of_blobs<false> chan{..., std::move(sock)}; // <-- A Unix domain socket pipe only.
chan.send_native_handle(...); // Compile error: Socket_stream_channel_of_blobs cannot transmit FDs.
chan.send_blob(...); // No problem.
Mqs_channel<false, Bipc_mq_handle> chan{..., std::move(out_q), std::move(in_q)}; // <-- A boost.interprocess-MQ pipe only.
chan.send_native_handle(...); // Compile error: MQs can certainly not transmit FDs.
chan.send_blob(...); // No problem.

All those subtleties about different types of pipes in a Channel bundle completely disappear when one deals with structured-data struc::Channels. They are a higher-layer abstraction; each will leverage whatever transport::Channel it adapts. In addition to handling capnp-encoded structured data and SHM-backed zero-copy, it also provides basics like request/response, request-method multiplexing, and a bit more. So let's get into that.

Transport (structured) (via struc::Channel)

While a Channel transports blobs and/or native handles, it is likely the Flow-IPC user will want to be able to transmit schema-based structured data, gaining the benefits of that approach including arbitrary data-structure complexity and forward/backward-compatibility. capnp (Cap'n Proto) is a leading third-party framework for schema-based structured data; ipc::transport's structured layer works by integrating with capnp.

Note

In this section we discuss a Flow-IPC-native structured-messaging technique (struc::Channel). A mature alternative is capnp-RPC (with Flow-IPC's zero-copy): See Transport (structured): capnp-RPC below.

To deal with structured data instead of mere blobs (though a schema-based structure can, of course, itself store blobs such as images), one simply constructs an ipc::transport::struc::Channel, feeding it an std::move()d already-opened Channel. This is called upgrading an unstructured Channel to a struc::Channel. A key template parameter to struc::Channel is a capnp-generated root schema class of the user's choice. This declares, at compile-time, what data structures (messages) one can transmit via that struc::Channel (and an identically-typed counterpart struc::Channel in the opposing process).

struc::Channel provides fundamental niceties for a structured-channel protocol:

• If I receive a message M of type X (from available types A, B, C, ... as allowed by the top-level capnp-union in the specified root schema), I ask struc::Channel to please invoke a particular handler F(M) (often given as a lambda).
• When I send a message M', I can optionally specify that it's a response to earlier in-message M.
• When I send a message M', I can optionally specify that I expect either up-to-1, or an arbitrary number of, responses to it. I ask that when such a response M is received, struc::Channel please invoke a particular handler F(M).

struc::Msg_out is a container-like data structure representing an out-message M'. struc::Msg_in is a receiver's read-only view into such a message M upon its receipt.

Synopsis: Transport (structured)

‍Like any Cap'n Proto user, to define your messages you will write a schema file which capnp will compile into C++ code with getters and setters. A refresher (or intro):

# A schema file defines your potential structures.
# ...
$Cxx.namespace("my_meta_app::capnp"); # Resulting generated classes will be in this C++ `namespace`.
 
struct CoolMsg
{
  # You can use absolutely any schema you would as a general capnp user... except:  We add *exactly one* requirement:
  # Your root `struct` must contain an anonymous union (we will use this to mux/demux in-messages to C++ callbacks).
 
  description @0 :Text;
  # You can, if desired, have arbitrary fields (like this one) outside the union, to be present in any message.
 
  union
  {
    # The union contains your possible messages (each of which can be of any type but usually a `struct`).
    sumRequest @1 :SumRequest;
    sumResponse @2 :SumResponse;
  }
}
struct SumRequest
{
  # Asks that receiver add up termsToAdd, multiply sum by constantFactor, and return result in
  # SumResponse.
  termsToAdd @0 :List(Int64);
  constantFactor @1 :Int64;
}
struct SumResponse
{
  result @0 :Int64;
}

You have already created an opened unstructured channel (transport::Channel raw_chan), most likely from a session. An example of transmitting structured data follows. The basic procedure:

1. Upgrade it to a structured channel (transport::struc::Channel).
2. Create a message. (struc::Channel::create_msg() is usually easiest. You can also construct explicitly for advanced needs.)
3. Use ipc::transport::struc::Channel APIs to issue and expect messages of various types. Request/response and other paradigms are supported.

Example: On one side, send a request.

using namespace ipc::transport::struc;
namespace ssn = ipc::session; // In this example, use non-zero-copy (internally, heap-based) transmission.
using Session = ssn::...;
 
// ...session+channel opening...
// raw_chan = opened channel from session.
 
// Upgrade it to a struc::Channel.  Specify your capnp-generated schema as template parameter.
Session::Structured_channel<my_meta_app::capnp::CoolMsg> // <-- Attention.
  chan{nullptr, std::move(raw_chan), // Eat it!  `raw_chan` becomes empty.
       Channel_base::S_SERIALIZE_VIA_HEAP, session.session_token()};
chan.start(...); // Error handler (omitted).
 
// Create a message -- no need to worry about capnp MessageBuilder work!  In our case it allocates in heap.
auto msg = chan.create_msg();
const auto body_root = msg.body_root();
// From here... back to vanilla capnp land!  To wit:
auto root = body_root->initSumRequest();
auto list = root.initTermsToAdd(3);
list.set(0, 11);
list.set(1, -22);
list.set(2, 33);
root.setConstantFactor(-2);
body_root->setDescription("(11 - 22 + 33) x -2 = -44");
 
// .sync_request(): send request; synchronously await response.
auto rsp = chan.sync_request(&msg);
// Note: The `msg` arg is a non-const pointer; `msg` shall be modified.  However this will not in fact modify the
// payload (mutated above) in `msg` but rather merely certain internal-use metadata info co-stored in the
// `Msg_out` object for perf reasons.
rsp = chan.sync_request(&msg, nullptr, chrono::seconds{5}, nullptr); // Or: with optional timeout.
 
// The other side had best know how to add stuff and multiply at the end!  Check the in-message.
assert(rsp->body_root().getSumResponse().getResult() == int64_t(-44));

On the other side: Expect request. Send response.

// Same deal as on the other side.
Session::Structured_channel<my_meta_app::capnp::CoolMsg>
  chan{nullptr, std::move(raw_chan), Channel_base::S_SERIALIZE_VIA_HEAP, session.session_token()};
chan.start(...);
 
chan.expect_msgs(my_meta_app::capnp::CoolMsg::Which::SUM_REQUEST, // Demux to handler on receipt of SumRequest union-choice.
                 [...](auto&& req) // Type is shared_ptr<M>, where M is: struc::Channel<...>::Msg_in, a/k/a struc::Msg_in<...>.
{
  // ...event loop brief detail omitted...
 
  const auto body_root = req->body_root();
 
  // Vanilla capnp work:
 
  auto root = body_root.getSumRequest();
  auto list = root.getTermsToAdd();
  int64_t sum = 0;
  const auto sum_func = [&](int64_t val) { sum += val; };
  std::for_each(list.begin(), list.end(), sum_func);
  sum *= root.getConstantFactor();
 
  std::cout << "Got request [" << *req << "] including description "
               "[" << body_root.getDescription() << "]; responding with result [" << sum << "].\n";
 
  // Respond.
 
  auto msg = chan.create_msg();
  msg.body_root()->initSumResponse().setResult(sum);
  msg.body_root()->setDescription("the sum you ordered, sir");
 
  chan.send(&msg, req.get()); // <-- originating_msg_or_null.  We are responding.
  // Note: As with sync_request() above, `msg` is modified (for perf reasons) but not the actual payload
  // as mutated-in just now.
 
  // ...event loop brief detail omitted...
});

Beyond that, any kind of code-snippet synopsis would be too wordy for this page; see Structured Message Transport (Structured Channels) for in-depth coverage. We'll summarize the capabilities here:

• Each out-message is either unsolicited (not responding to any particular in-message) or a response (responds to one specific in-message). Every send-method of ipc::transport::struc::Channel (and struc::sync_io::Channel) takes an "originating message" pointer: simply supply nullptr for unsolicited messages or a Msg_in for responses.
  • This applies to: .send(), .async_request(), .sync_request().
• Each in-message is therefore either unsolicited or a response, naturally.
  • If you expect a response, this is indicated by issuing a request: use .async_request() (response arrives asynchronously) or .sync_request() (as in the above example).
    • .async_request() can be one-off or indefinite-lifetime, depending on a certain argument. In the latter case 0+ responses can arrive, and you can explicitly unregister the expectation for more responses using .undo_expect_responses().
    • .sync_request() is by definition one-off: It waits for 1 response exactly.
      • You can optionally set a timeout. In many cases it is not necessary, as you can count on a quick (non-blocking) call. Just ensure you have a handler set up on the opposing side, before you begin issuing these, to avoid blocking.
  • If you expect an unsolicited message, call .expect_msg() or .expect_msgs(); you'll need to supply the union-selector enumeration value (SUM_REQUEST in the above example) to demux to that message type's specific handler.
    • Use .expect_msgs() (as in the above example) if you expect an arbitrary number of such unsolicited messages. Use .undo_expect_msgs() if desired later.
    • Use .expect_msg() if you expect exactly one such unsolicited message.
    • An unsolicited message that does not currently have a registered handler (outstanding .expect_msg*() call) is not dropped. It is cached. If a matching .expect_msg*() is subsequently called, its handler is immediately fed with any applicable cached message(s).
      • There is no caching for unexpected responses (response without request). .set_unexpected_response_handler() and .set_remote_unexpected_response_handler() can detect such events, but you can also design your code/protocol to not need them. In general response-without-request is dropped: it is not a channel-hosing condition.
• Each out-message is either a notification (expects no response) or a request.
  • .send() is for notifications. For example the SumResponse sent in the example is a notification.
  • .*sync_request() is for requests.

For Cap'n Proto enthusiasts, ipc::transport::struc::Msg_out provides a couple ways of structuring your message-filling in addition to the vanilla way shown in the example above.

• Msg_out::orphanage() lets one operate a capnp orphanage. You can build things there in bottom-up fashion and then ...adopt....() them into the root message.
• Construct Msg_out explicitly, by giving it an existing capnp::MessageBuilder. This allows for completely arbitrary Cap'n Proto mutation work, without having to use any orphanage.

For advanced needs, you can replace some of the default machinery in .create_msg(). For example you can obtain backing memory for out- and in-messages from a mem-pool, or (if performing non-concurrent work) reuse a single block of memory. (See Struct_builder and Struct_reader concepts.)

Transport (structured): Zero-copy

In the example just above, the request out-message is copied into the IPC transport (e.g., Unix domain socket) internally on one side; and copied again out of the transport on the receiver side. The in-message's backing memory is (internally) allocated in heap; as is the original out-message's. It need not be copied at all. This is easily done in Flow-IPC.

Say I prepare a message M': I construct it (struc::Msg_out), I mutate it via its capnp-generated mutators (as the required backing memory is gradually allocated behind the scenes), and then I struc::Channel::send() it. The receiver receives it as struc::Msg_in M subsequently and accesses its various parts via capnp-generated accessors. End-to-end zero-copy means that the backing memory allocated for M' is never copied into any low-level transport. Instead only some tiny, constant-sized handle is (internally) copied into and out of the transport – and the accessors of M are directly reading the memory in M'.

The benefit: message size stops mattering, perf-wise. Notably, also, it is possible to subsequently modify M' and send it (and receive it) again. The backing memory is freed once M' and all in-message views like M have been destroyed across all processes. (The Ms are made available exclusively via shared_ptr.)

Zero-copy is optional. It is quite easy to enable. Internally it is accomplished by using SHM; but in terms of the public API, ~all you have to do is – when writing the code for opening your ipc::session::Session – simply choose a SHM-backed session type. That is:

• Zero-copy (a/k/a SHM-backed): Use "ipc::session::shm::...::Client_session" and "ipc::session::shm::...::Session_server".
• Non-zero-copy: Use ipc::session::Client_session and ipc::session::Session_server.

Whichever you choose, their APIs are identical. In particular, suppose you chose Session type S; then S::Structured_channel<Your_capnp_schema> is the alias for the proper SHM-backed-or-not struc::Channel type.

At no point do you have to worry about naming a SHM pool, removing it from the file-system on cleanup (including following a crash), and so on. All is handled internally. Simply choose a SHM-backed Session type on each side at compile-time.

Synopsis: Transport (structured): Enabling zero-copy end-to-end

‍Consider the structured-transmission example just above. It is not zero-copy end-to-end. Here is all you have to do to make it zero-copy end-to-end. Notice the tiny change and zero mentions of any SHM details.

// ...everything same...
 
using Session = ssn::...;
// --- CHANGE TO --v
using Session = ssn::shm::classic::...;
 
// ...everything same...
 
Session::Structured_channel<my_meta_app::capnp::CoolMsg>
  chan{nullptr, std::move(raw_chan),
       Channel_base::S_SERIALIZE_VIA_HEAP, session.session_token()}; // <-- CHANGED LINE.
// --- CHANGE TO --v
Session::Structured_channel<my_meta_app::capnp::CoolMsg>
  chan{nullptr, std::move(raw_chan),
       Channel_base::S_SERIALIZE_VIA_SESSION_SHM, &session}; // <-- CHANGED LINE.
 
// ...everything same...

Variation: switch SHM-providers:

// As written above, when allocation is necessary internally, SHM-classic provider will use a classic one-segment SHM setup,
// with a simple allocation algorithm provided by boost.interprocess.
using Session = ssn::shm::classic::...;
// --- CHANGE TO --v
using Session = ssn::shm::arena_lend::jemalloc::...;
// Now SHM-jemalloc provider is used instead; it will internally use multiple SHM segments as needed, driven by
// commercial-grade `jemalloc` allocation algorithm.  Otherwise example code remains unchanged.

Variation: change max lifetime of messages created via struc::Channel::create_msg():

// As written above, the maximum lifetime of a `Msg_out` = as long as the session remains open.
// This is often clean and sufficient.
Session::Structured_channel<...> chan{...,
                                      Channel_base::S_SERIALIZE_VIA_SESSION_SHM, ...}; // <-- CHANGED LINE.
// --- CHANGE TO --v
// Now the message can live *past* the session: until the Session_server is destroyed, meaning
// your process accepting incoming sessions exits.  For example, suppose your session-server accepts
// (sessions) from many session-client processes of one application over its lifetime; session-client A sends
// a PutCache message containing a large file's contents to session-server which memorizes it (to serve later);
// then session-client A closes session and exits; the message continues to exist, and the memory-cached file
// can be served to session-clients B, C, ....
Session::Structured_channel<...> chan{...,
                                      Channel_base::S_SERIALIZE_VIA_APP_SHM, ...}; // <-- CHANGED LINE.

In our example in particular, there is simply no difference between using heap-backed allocation and SHM-backed allocation. Just change a couple lines, and you get zero-copy. The messages being exchanged can be utterly huge, with zero perf penalty.

That said, for more complex use-cases, there are natural differences in what one can do when comparing heap-backed versus SHM-backed allocation. For example, the out-message can be modified after sending it, and this change will be reflected for any process(es) holding the resulting in-message(s). (Of course you'll need to arrange some kind of synchronization of access in that case.)

In any case – whether heap-backed or SHM-backed – an out-message can be sent or re-sent over any channel and even in multiple sessions, depending on the details of its backing memory allocation.

Transport (structured): capnp-RPC

struc::Channel is one way to exchange capnp-structured data. The other is Cap'n Proto's own Remote Procedure Call layer (capnp-RPC): interfaces implemented by one process and called by another, with promise pipelining and more (see capnp-RPC docs). Flow-IPC makes capnp-RPC zero-copy-enabled. It supplies a drop-in replacement for capnp's TwoPartyVatNetwork, plus counterparts of capnp's other connection-setup helpers. The rest of your capnp-RPC code stays the same. Instead of socket addresses, you specify the opposing process using the same Apps as with sessions.

Synopsis: Zero-copy capnp-RPC (the easy way)

‍Counterparts of capnp's EzRpcServer and EzRpcClient; using capnp's Calculator sample:

// Choose the SHM-provider.  (The App objects are as in the sessions synopsis above.)
using Session_server = ipc::session::shm::classic::rpc::Session_server<>;
using Client_session = ipc::session::shm::classic::rpc::Client_session<>;
 
// Server application:
ipc::transport::struc::shm::rpc::Ez_rpc_server<Session_server>
  server{nullptr,
         [](auto&&...) { return kj::heap<CalculatorImpl>(); }, // Bootstrap interface impl for each client.
         APP_A_AS_SRV, MASTER_APPS_AS_CLI};
kj::NEVER_DONE.wait(*(server.get_wait_scope())); // Accept and serve connections forever.
 
// Client application:
ipc::transport::struc::shm::rpc::Ez_rpc_client<Client_session> client{nullptr, APP_B_AS_CLI, APP_A_AS_SRV};
Calculator::Client calculator = client.get_main<Calculator>();
// ...From here: ordinary capnp-RPC code.

See Zero-copy capnp-RPC for the full story, including the counterparts of RpcSystem-level vanilla capnp-RPC setup.

Direct storage in SHared Memory (SHM)

This topic is optional: One can get zero-copy by exclusively keeping IPC-communicated data structures in your struc::Channel-transmitted message schema of choice. Internally, we use our own SHM capabilities to accomplish this, invisibly to the API user. For some use cases, if you can keep your shared data structures in a capnp-schema-backed object, then it is best to do so. That said:

Some applications may prefer to have the two (or more) conversant processes cooperate by reading/writing directly to a straight C++ struct rather than a capnp-backed data structure. Flow-IPC provides this ability out-of-the-box. This means the user need not figure out a number of very hairy topics such as how to develop and use a SHM-friendly allocator. We summarize this here.

If and only if you used a SHM-enabled ipc::session::Session implementation (as noted in the preceding section – for example ipc::session::shm::classic::Client_session), then upon opening such a session, you immediately have access to 2 SHM arenas.

• Session::session_shm() is an arena pertaining to that session (pair of processes).
• Session::app_shm() is an arena pertaining to multiple sessions; so it continues to persist across all sessions, until the big-daddy Session_server is destroyed (and thus no further sessions are possible to open).

Each of these returns an Arena.

Arena::construct<T>(...) returns shared_ptr<T> which points to a newly constructed-in-SHM-arena object of type T. The sender prepares object *x of type T, then calls Session::lend_object(x) which returns a tiny payload to transmit over IPC (typically, but not necessarily, via a struc::Channel as part of a capnp-encoded message). The receiver then invokes auto x = Session::borrow_object(), passing in that aforementioned tiny payload. This recovers *x of type T in the receiver where it can be read like any C/C++ structure – because it is one. The backing RAM is auto-returned to the arena once all these handles (in this example x on each side) (and other derived shared_ptrs in their respective shared-groups – e.g., obtained via auto x2 = x) are destroyed. In a sense x is part of a cross-process shared_ptr group.

The real power comes from the possibilities for what type T can be. T can be any combination (recursively speaking) of: the basic scalars (int, float, etc.); structs; and STL-compliant containers. So it can be, say, a vector of unordered_maps mapping from basic_string to struct S { ... }, where ... itself stores anything equally arbitrarily complex. The STL-compliant container types must merely specify the allocator ipc::shm::stl::Stateless_allocator<Arena>. We've provided the allocator implementation, so you need not worry about that stuff – just use it.

You can also directly store pointers in T, as long as you use Stateless_allocator<Arena>::Pointer<P> (not raw P*); though we'd informally recommend against it for the maintainability of your own algorithms. That is what the allocator concept in STL is for after all. (This advice is independent of Flow-IPC; just a general opinion.)

Further capabilities are outside our scope here; but the main point is: At a minimum:

• if you simply used a SHM-backed session type when setting up IPC, immediately available to you are arenas in which to allocate objects and share them between the processes involved; and
• the object type can be an arbitrarily complex combo of C++ plain-old data-types (PODs) and STL-compliant container types, as long as you use the Flow-IPC-supplied allocator for the latter.

Synopsis: Native STL-compliant data structures directly in SHared Memory: construct, transmit, access

‍A SHM-compatible structure being declared and an instance constructed and modified:

// Define an allocator alias for convenience.
// (For plain-old-datatypes (e.g., an `int` or `bool` or `struct` thereof or combinations thereof)
// there is no need to use this.)
template<typename T>
using Shm_allocator = Session::Allocator<T>;
 
// Example complex structure.
struct Widget
{
  struct Node
  {
    int m_int;
    // Normally would be just `vector<float> m_float_vec`.
    boost::container::vector<float, Shm_allocator<float>> m_float_vec;
  };
  // Normally would be just `string`.
  using String = boost::container::basic_string<char, std::char_traits<char>, Shm_allocator<char>>;
 
  bool m_flag;
  String m_str;
  // Normally would be just unordered_map<string, Node> m_str_to_node_map.
  boost::unordered_map<String, Node, std::hash<String>, std::equal_to<String>,
                       Shm_allocator<std::pair<const String, Node>>
    m_str_to_node_map;
};
 
// Construct it directly in SHM.  (Default ctor used here; you can use any ctor.)
auto x = session.session_shm()->construct<Widget>();
 
// Modifications not involving more allocation are done as-normal.
x->m_flag = true;
x->m_str[0] = 'X'; // Assumes, of course, that by this point: x->m_str.size() > 0.
 
// If allocation is required, apply a thread-local Activator.  Can frame all modification code in such a block.
using Activator = ipc::shm::stl::Arena_activator<Session::Arena>;
 
auto y = session.app_shm()->construct<Widget>();
{
  Activator ctx{session.app_shm()}; // Same arena as `y`'s.
 
  y->m_flag = true; // `ctx` is active but has zero effect on this (harmless and zero perf impact).
  y->m_str.resize(128); // `ctx` is active and will cause proper arena to be used for allocation.
  y->m_str[0] = 'X'; // `ctx` again has no effect and carries no perf penalty for this.
}

Example of transmitting a SHM-backed native data structure to another process follows. We transmit a handle through a capnp structured message here, but it can be done using any IPC mechanism whatsoever; even (e.g.) a file.

Schema which includes a native-object-in-SHM field:

# ...
using Common = import "/ipc/transport/struc/shm/schema/common.capnp";
using ShmHandle = Common.ShmHandle; # It's just a blob holder really.
# ...
struct SomeMsg
{
  # ...
  widgetHandle @7 :ShmHandle;
}

Owner/lender process transmitting such a message:

auto x = session.session_shm()->construct<Widget>();
// ...Fill out *x as above....
 
auto msg = chan.create_msg();
// Ready the storage inside the out-message.
auto widget_handle_root = msg.body_root()->initSomeMsg().initWidgetHandle(); // Note `widgetHandle: ShmHandle` in schema.
// Perform the lend_object() step and load the result (a little blob) into the out-message.
shm::capnp_set_lent_shm_handle(&widget_handle_root, session.lend_object(x));
// IPC-transmit it via struc::Channel.
chan.send(&msg);
// Note: Again, `msg` is modified (for perf reasons) but not the payload (mutated-in just above) itself.

Borrower process receiving such a message and accessing the native data structure Widget:

flow::util::Blob_sans_log_context lend_blob;
shm::capnp_get_shm_handle_to_borrow(msg->body_root().getSomeMsg().getWidgetHandle(), &lend_blob);
 
auto x = session.borrow_object<Widget>(lend_blob);
FLOW_LOG_INFO("Hey, let's read inside SHM after receiving SHM-handle: [" << x->m_flag << "].");
// We can write to it too, in most cases.
x->m_str[0] = 'Y';
// (Can modify with allocations too, as long as you use an Activator context again, same as in owner process above.)

When one gets into this relatively advanced area, there are differences and trade-offs between the available SHM-providers. If your alias at the very top is using ssn = ipc::session::shm::classic, then you get total read/write abilities and maximized object lifetime. That is the SHM-classic provider. Instead, if you choose SHM-jemalloc (using ssn = ipc::session::shm::arena_lend::jemalloc), then the borrower (receiver) process can only read, and there are some subtle differences in max object lifetime. (The latter applies only if you need objects to live beyond a given Session.) However, you get in return: superior safety characteristics; and the commercial-grade jemalloc allocation algorithm (thread-caching, fragmentation avoidance, and more).

SHM without sessions: SHM-classic standalone

All of the above obtains arenas from a Session. That is the easy way, as ipc::session names, creates, and cleans up the underlying SHM pools for you; plus the means of transmitting the lend_object()-returned mini-blobs are right there (the session-generated Channel(s)). However the SHM-classic provider is perfectly usable on its own, which may be convenient for quick prototyping, say. SHM-classic's arena type, ipc::shm::classic::Pool_arena, is a thin wrapper around a single named SHM pool (in the spirit of bipc::managed_shared_memory), with the very same construct-lend-borrow API seen above. You name the pool, open it in each process, and transmit the lend-blobs by any means you like. You are also responsible for removing the pool when done.

‍Create, lend, borrow, clean up – SHM-classic, no session:

using ipc::shm::classic::Pool_arena; // (Same type as: `ipc::session::shm::classic::Session::Arena`.)
using ipc::util::Shared_name;
const auto POOL_NAME = Shared_name::ct("..."); // Some name both processes agree on.
 
// Process A: create the pool.  The size can be huge: RAM is committed only as pages are touched
// (see Pool_arena docs).
Pool_arena arena_a{&logger, POOL_NAME, ipc::util::CREATE_ONLY, size_t(1) << 30,
                   ipc::util::shared_resource_permissions(ipc::util::Permissions_level::S_GROUP_ACCESS)};
auto x = arena_a.construct<Widget>(); // Widget declared with Pool_arena::Allocator; fill it out with
                                      // a Pool_arena::Activator active, as above.
const auto lend_blob = arena_a.lend_object(x); // Transmit `lend_blob` to process B somehow.
 
// Process B: open the same pool; borrow the object.
Pool_arena arena_b{&logger, POOL_NAME, ipc::util::OPEN_ONLY};
auto x = arena_b.borrow_object<Widget>(lend_blob);
// ...Read/write *x....
 
// When everyone is done with the pool:
Pool_arena::remove_persistent(&logger, POOL_NAME);
 
// A Pool_arena is, by default, sparse (RAM committed as pages are touched).  It's possible, instead, to commit the
// full size anytime.  If arena use might exceed an OS SHM-use limit, this is the civilized way of catching it.
// (The trade-off: No longer advisable to just specify a huge pool size, as that would be real RAM use now.)
Pool_arena arena_c{..., ipc::util::CREATE_ONLY, 2 * 1024 * 1024, ...}; // 2Mi bytes...
arena_c.commit(); // ...taken immediately from general RAM.
// If limit exceeded: Threw exception E.  E.code() and E.what() then encode an OS no-space-left error.

SHM-jemalloc, too, can be used without ipc::session, but the setup is much more involved: One needs an ipc::shm::arena_lend::jemalloc::Ipc_arena (owner side) plus an ipc::session::shm::arena_lend::jemalloc::Shm_session per opposing process, the latter requiring an established channel for its internal messaging. See those classes' doc headers.

Safety

Flow-IPC focuses on IPC safety. Safety and Permissions page gets into the topic in detail. Here we'll point out merely one specific mechanism which might be of immediate interest. It shows that – while at various lower layers permissions can be specified in fine-grained fashion – at the top layer, ipc::session, we let you decide one of 3 simple safety approaches matching your setup, and Flow-IPC will take care of the details accordingly. To wit:

Refer to the very top of Synopsis: Sessions above. When defining our simple Server_app structure that describes the application responsible for accepting sessions (IPC conversations/contexts), we chose:

‍The Server_app definition from that synopsis:

const App APP_A{ "a", "/bin/app_a.exec", USER_ID_A, GROUP_ID };
const Server_app APP_A_AS_SRV{ APP_A,
                               // ...
                               ipc::util::Permissions_level::S_GROUP_ACCESS }; // <-- Attention.

Notice the GROUP_ACCESS choice made there. In fact we could have chosen any one of these:

• S_UNRESTRICTED:
  • This will ensure that the relevant resources will never fail to be accessed by the relevant processes due to a permissions error. (Internally, producer-consumer items like PID files shall be 0644; while symmetrically accessed items like SHM pools and MQs shall be 0666.)
• S_USER_ACCESS:
  • This means access to a resource may fail if the accessing application is running as a different user than the writing application, even if they are in the same group. (Internally items shall have permissions mode 0600.)
• S_GROUP_ACCESS:
  • This means access to a resource may fail if the accessing application is running as a different group than the writing application, but it will work if they're in the same group but are running as different users. (Internally, producer-consumer items like PID files shall be 0640; while symmetrically accessed items like SHM pools and MQs shall be 0660.)

So: GROUP_ACCESS means your various apps should run as different users but in the same group. USER_ACCESS means the apps should run all as one user. UNRESTRICTED means it does not matter.

Stats

Flow-IPC keeps extensive statistics: counts, gauges, high-water marks, histograms – of messages, bytes, requests, allocations, SHM-pools, and much more – throughout its layers. Consuming them is straightforward:

• Each stats getter emits a plain struct of scalars (a stat-set), a snapshot as of the call. Read its members; or print the whole thing via ostream << flow::util::stat::print(the_stat_set).
• Next to each getter there is a matching *_reset() method. In some cases a *_reset() method affects data emitted by more than one getter.
• Some areas additionally offer a bundle that gathers several stat-sets (and related info) into one object, printable via ostream << in one shot (and resettable, where applicable, via the individual *_reset()s).

Every stat-set member is documented in the Reference. We do not re-document them in this Manual; but we do point out that those docs are a public, reader-friendly, source of information on how things work internally.

‍Printing and resetting stats, and printing bundles:

// Any stat-set: print it; reset it.
FLOW_LOG_INFO("Channel stats: " << flow::util::stat::print(cool_channel.stats()));
cool_channel.stats_reset();
// Bundles print in one shot.
FLOW_LOG_INFO("Session (internal master channel) info: " << *session.info_collector());
// Bundles have simple format knobs; e.g., skip the (huge, SHM-jemalloc-only) native jemalloc dump.
Session::Arena::Info_dump dump;
dump.m_fmt.m_verbose = false;
session.session_shm()->info_dump(&dump);
FLOW_LOG_INFO("SHM-arena info:\n" << dump << '.');

Where to find them (each chapter on the relevant topic has a short Stats section with the full list):

Area	Stats getters (each with *_reset())	Bundle
Unstructured transport (Native_socket_stream, Blob_stream_mq_sender/receiver, transport::Channel)	blob_send_stats(), blob_receive_stats(), native_handle_send_stats(), native_handle_receive_stats()	–
Structured transport: struc::[sync_io::]Channel	ipc::transport::struc::Channel::stats(), sync_io::Channel::stats()	–
Structured transport: serializers (process-wide)	via flow::util::stat::Global_stats singletons	ipc::transport::struc::stat::heap_serializer_info_dump(), ipc::transport::struc::shm::stat::serializer_info_dump()
Sessions (session master channel)	via ipc::session::Session::info_collector()	*session.info_collector()
SHM-classic arena	ipc::shm::classic::Pool_arena::arena_stats(), local_stats()	ipc::shm::classic::Pool_arena::info_dump()
SHM-jemalloc	ipc::shm::arena_lend::jemalloc::Ipc_arena (pool_stats(), memory_manager_stats(), sharded_stats(), and more); ipc::session::shm::arena_lend::jemalloc::Shm_session (borrower_pool_stats(), and more)	Ipc_arena::info_dump(), Shm_session::info_dump()
capnp-RPC	(none of its own: use those of the underlying session and its channels/arenas)	–

---

Flow module flow::util::stat

If you want to post-process rather than merely read/print – e.g., aggregate stat-sets from multiple like objects – see the flow::util::stat namespace doc header in Flow.

Furthermore, that general Flow module is available for your applications to collect their own stats. The basic premise behind the stats module: the outputs are straightforward structs that shall be easily printable; declaring and filling them out, at minimal boiler-plate and perf costs, is a deep topic. Flow-IPC exercises much of this, as there is a significant variety of modules, with various concurrency profiles, to cover.

---

Next let's get to the nitty-gritty of actually using this library: Prerequisites and Setup.

---

MANUAL NAVIGATION: Preceding Page - Next Page - Table of Contents - Reference