|
Flow-IPC 1.0.0
Flow-IPC project: Full implementation reference.
|
This page explains how to send and receive structured messages, including requests, over a channel – for many users the essential purpose of the library. (Or go back to preceding page: Multi-split Universes. Sessions: Opening Channels would also be a natural immediate pre-requisite for the current page.)
To restate from Sessions: Opening Channels – a channel is a bundling of the peer 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 handle and a similar incoming-MQ handle; or conceptually similar SHM-backed MQs (boost.ipc MQs). A peer Channel can be upgraded to an ipc::transport::struc::Channel (or sync_io-pattern counterpart, ipc::transport::struc::sync_io::Channel) in order to represent a structured, arbitrarily-sized datum per message as opposed to a mere (limited-size) blob.
Put more abstractly, however, a channel is, well, a channel of reliable bidirectional messaging. That said, if one takes a look at its API – knowing nothing else – it will be obvious it transmits:
Each message contains either exactly one blob (whose size is limited by the compile-time-chosen low-level transport), exactly one native handle, or both.
Handles aside, typically it is not convenient to manually represent logical data as binary blobs. The size is limited and transport-dependent for one; but that aside it is simply inconvenient and potentially slow to pack data in this fashion. For this reason serialization frameworks exist, and we certainly had no desire to reinvent the wheel. We simply chose the best one – capnp (Cap'n Proto) – and built around its API. The result is: ipc::transport::struc::Channel (and/or its sync_io counterpart struc::sync_io::Channel).
The steps to work with a struc::Channel are as follows.
Channel. This is explained in-depth in Sessions: Opening Channels. (One can also open a channel manually without ipc::session. We will discuss this, typically advanced or optional ability, in later pages of this Manual.) The result shall be an object whose type is an instance of ipc::transport::Channel class template, or possibly a sub-class thereof. (In the latter case the sub-class adds no data on top of the Channel – only a small convenience API whose significance is low in our context here.)Channel, upgrade it to a struc::Channel object. This is done via C++ move semantics: The new struc::Channel constructor takes a Channel&& reference, thus taking over the guts of the unstructured Channel, while the original Channel object becomes blank – as-if default-constructed.struc::Channel.Channel is still available for limited needs, most notably ipc::transport::Channel::auto_ping() and ipc::transport::Channel::idle_timer_run().struc::Channel API directly, exclusively.Each side of the conversation follows the above procedure.
Let's discuss step 2 of this procedure: upgrading a Channel to struc::Channel.
Channel to struc::Channel By the time you've got your unstructured Channel, you'd already made decisions as to its underlying low-level transport(s). (If you used ipc::session, as recommended, you did this using the template knobs of Session_server and Client_session.) These details, roughly speaking, will not matter when using the upgraded struc::Channel. There are only a couple of things you must ensure about the underlying transports:
true. (Tweak the ipc::session compile-time knobs as needed (Sessions: Opening Channels). Specifically set the S_TRANSMIT_NATIVE_HANDLES template parameter to true.)struc::Channel, then the max-blob-size enforced by the Channel must be large enough to hold the serialization of any message you choose to send.So let's say you have your Channel named unstructured_channel, most easily obtained via ipc::session. (If you're the type of person to mind the details – such a channel from ipc::session will always be sync_io-core-bearing, meaning it won't have started background threads and is generally lighter-weight.) Do not mutate it in any way. Instead feed it to a constructor of your new struc::Channel.
That's easy enough in and of itself, but you'll need to make a couple of decisions first as to the type of this struc::Channel: it is, also, a template. In this page we're going with the highest-performance, most-mainstream scenario, so we can skip explaining various complexities (leave it to Structured Message Transport: Messages As Long-lived Data Structures / Advanced Topics). All ipc::session::Session concept impls supply a using struc::Channel alias. Thus, as shown in Sessions: Setting Up an IPC Context, you'll have a Session type alias ready. Then:
A very central decision is the choice of Message_body. This is the schema of the messages (native handles aside) you intend to transmit back and forth. (Each side must use the same Message_body, or ones that are bit-compatible per capnp requirements.) The design of the schema and the generation of the required header and source-code files are well outside our scope: please see capnp schema docs and capnp tool docs. We shall assume a straightforward schema like this for our example:
Once you've run this through capnp, you'll have a class my_meta_app::capnp::CoolMsg available – notably with nested classes CoolMsg::Reader and CoolMsg::Builder, for reading and mutating such messages respectively. (Please refer to capnp C++ docs for details.)
Therefore your specific struc::Channel type is easiest to express as follows:
You've got the type; and you've got the (so-far-unused) unstructured Channel unstructured_channel; and you've got the ipc::session::Session used to open that unstructured_channel. Now it's time to upgrade it into a Cool_structured_channel. Continuing our mainstream-oriented example here's how:
The main thing happening here is that 2nd constructor arg: cool_channel empties unstructured_channel and takes over its previous contents. From now on you will be controlling the underlying channel by operating its subsuming struc::Channel. As for the other arguments:
&session is straightforward enough: That's the session used to open the channel. Internally it's used for SHM-backing setup and a certain safety mechanism. Bottom line is you need not worry about it. (Structured Message Transport: Messages As Long-lived Data Structures / Advanced Topics gets into all that.)S_SERIALIZE_VIA_SESSION_SHM has to do with the lifetime of the SHM-backing for messages. For now don't worry about it. Just know that, as written, a given message object cannot survive the end of session. (There are ways to exceed it; once again that's covered in Structured Message Transport: Messages As Long-lived Data Structures / Advanced Topics.)Before we get to sending/receiving messages there's one more – optional – thing you may (or may not) wish to get out of the way. The mutually-complementary auto-ping and idle-timer features of Channel are available at this stage. These are discussed in detail in Transport Core Layer: Transmitting Unstructured Data, but the short version is: You may wish for your channel to self-destruct if the other side does not send a message with at least a certain frequency: idle timer. You may, further, wish for (invisible) messages (called auto-pings) to be sent as a form of keep-alive. (ipc::session leverages both features, so that if the other side appears dead due to not even sending auto-pings, the ipc::session::Session will report an error, at which point you could close all your related channels. Therefore, if you're using ipc::session, auto-pinging via your own channels may be of little use. However the idle-timer may still be useful by itself depending on your protocol.) Long story short, here is how to do it:
struc::Channel) APIs. As all over this manual our mainline discussion and examples are written in terms of the async-I/O API. If you wish to use the sync_io alternative (see the aforementioned page for reasons why) then use ipc::transport::struc::sync_io::Channel instead of ipc::transport::struc::Channel. It is, like all APIs, fully documented in the Reference section of the manual (just click the links in this paragraph to go to the relevant reference pages). The basic things you need to know are:sync_io type is obtained most easily by using the ipc::transport::struc::Channel::Sync_io_obj alias. So you could do: using Cool_structured_channel = Structured_channel_t<my_meta_app::capnp::CoolMsg>::Sync_io_obj.Cool_structured_channel use the exact same constructor form you would have used with an async-I/O struc::Channel. Therefore the code above constructing the cool_channel would be identical. (Note that both struc::Channels' constructors take the same sync_io-core-bearing Channel type object to subsume. It does not matter that in one case the result is an async-I/O thing versus a sync_io thing.)sync_io API. It is similar, but not identical, to the async-I/O one.struc::Channel to transmit Naturally you'll need something to send, and therefore receive, before you can do that. You'll need to create an out-message as per ipc::transport::struc::Msg_out API. Let's get into that.
struc::Msg_out is, in essence, not too much different from a container. You declare it; you construct it (on the stack if desired); you mutate it; you destroy it. struc::Msg_outs are not copyable, but they are (cheaply) movable (move-constructible, move-assignable). While they can be constructed explicitly, it's by far easiest to leverage your struc::Channel – in our case Cool_structured_channel:
Assuming you are familiar with the capnp-generated API, it is straightforward to mutate the (currently blank) msg out-message. The only constraint in terms of the structured content of msg is that (as noted before) at the root level there is a mandatory anonymous union. So let's say we want to send a HelloThere message. Then we could write:
Builder, .getX() and .setX() are lightning-fast, like accessing struct members directly – but only when X is of a scalar type. Compound types, where .getX() returns not a native type but another Builder, need to perform some pointer checking and are slower. Therefore, if you plan to .getX() and then .setA() and .setB() (and similar) on that X, you should save the result (auto x = ....getX();); then mutate via the saved result (x.setA(...); x.setB(...)).X be a compound field, particularly List, Text (string/list-of-characters), Data (list-of-bytes). It is usually, to begin, null; you must perform .initX(size_t n) or equivalent to initialize it (fill it with n zeroes). However, suppose you are mutating a message, such as a previously-sent message, and wish to modify the X. If the new value might have the same length (which is common), the correct thing to do is: check .hasX(); if true, modify .getX(); but if false then .initX(n) (as X was null after all). Performing .initX() on an already-.initX()ed value works but has a nasty invisible effect: the existing datum continues taking space in the serialization; the new datum takes a new chunk of the serialization segments (and might even cause the allocation of a new segment if needed). As of this writing capnp does not reuse such orphaned space. If the new n equals the old n, this is a straight waste of RAM; and can lead to pathologically growing memory leaks if done many times.n is different from the preceding, then there is no choice but to re-.initX(). A list/blob/string's size cannot be modified in capnp. It is best to avoid any situation where the n would change; try to design your protocol differently.)ostream << to pretty-print a struc::Msg_out, without newlines/indentation and truncated as needed to a reasonable length. For fully indented pretty-printing you may use capnp::prettyPrint(msg.body_root()->asReader()).flatten().cStr(). (Be wary of the perf cost of such an operation, especially for large messages. Though if done within a FLOW_LOG_*() no evaluation occurs, unless the the log-level check passes.)More advanced capnp-leveraging operations are possible. We omit detailed discussion here. Briefly:
Orphanage via msg.orphanage() (see ipc::transport::struc::Msg_out::orphanage()). With this you can make objects of any type(s) whatsoever and then adopt...() them in a bottom-up fashion into the ultimate msg.body_root() (which is, rigidly, a CoolMsg::Builder). See capnp docs.capnp::MessageBuilder – without any particular set root schema – and then later load it into a struc::Msg_out by using an alternate constructor form. In other words one can use a capnp::MessageBuilder as a heap for arbitrary work. Just ensure that what you're ultimately loading into the struc::Msg_out has most recently been .initRoot<M>()ed, where M matches the Message_body type of your struc::Channel. (So in our example M is CoolMsg.)Lastly, assuming your underlying Cool_structured_channel::Owned_channel is capable of it, any out-message may be paired with a native handle like this:
So you've loaded up your struc::Channel::Msg_out a/k/a struc::Msg_out. Now you'll presumably want to send it. The Msg_out is a message, while a message actually sent via channel (and received by the opposing side) is a message instance. Therefore the act of invoking one of the send methods of a struc::Channel (.send(), .async_request(), .sync_request()) creates a message instance from a message – which the opposing side obtains as an ipc::transport::struc::Msg_in (but more on this later). All properties of a message instance are contained either in the message itself or in the arguments to the chosen struc::Channel send method; they cannot be changed subsequently.
A message is a message; but a message instance is one of 2 things:
Either type of message – whether notification or response – can itself be in response to an earlier-received in-message, or not in response to anything.
.send() or .*sync_request().Thus .send() is the simplest method to demonstrate:
That's it. A key point: .send() – as well as .async_request() –
That last point is unusual but intentionally nice. You never need to worry about a send failing due to a clogged-up low-level transport. (How this is accomplished is discussed in Transport Core Layer: Transmitting Unstructured Data – but it's an advanced topic, and you can usually just take it for granted.)
To send a request use .async_request() or .sync_request(). The latter blocks until the response arrives and returns the response on success, synchronously.
.sync_request() is, indeed, a rare blocking method in Flow-IPC. However in many (certainly not all) situations one can be reasonably assured that a properly behaving opposing process is currently expecting a message of a certain type and will choose to immediately .send() (et al) a response. IPC is meant to be, and is, low-latency – therefore in such a setup one can reasonably expect a response to arrive with an RTT (round-trip time) of microseconds. Therefore, to simplify code, a .sync_request() can and should be used where possible as a "non-blocking blocking" call. (For worried souls, an optional timeout argument is available.) Otherwise use .async_request() (explained a bit further below).Example:
.sync_request() blocks up to the timeout; but if a channel-hosing error occurs during its operation it will return immediately emitting the error (which may or may not be specifically related to the send attempt itself). (We discuss error handling in more detail below.) However the specific timeout error is not fatal: ipc::transport::error::Code::S_TIMEOUT..sync_request() is not special in terms of thread safety: You may not invoke other methods concurrently on the same struc::Channel, until it has returned..sync_request() is by definition one-off: If the opposing side sends 2 or more responses, all but the first will be ignored (logging aside), except unexpected-response handler(s) may fire (more on this below). C/f .async_request() which allows one-off and indefinite-lifetime requests alike.That brings us to .async_request(), the other way of sending a message (turning it into a message instance) while expecting response(s). Depending on the situation it may be less or more mainstream than .sync_request(). An async-request is as follows:
.async_request(), one specifies whether:In any case, if the opposing side sends a response that is not expected at a given time, it is ignored (logging aside), except unexpected-response handler(s) may fire (more on this below).
We've already demonstrated (more or less) dealing with a one-off request, so for this example let's issue an indefinite-lifetime one.
This following code assumes an async-I/O event loop setup introduced here – in this case a single thread called thread W, onto which further tasks may be posted using a post() function.
sync_io-pattern ipc::transport::struc::sync_io::Channel::async_request() conceptually is quite similar to the async-I/O one shown above, and the code will likely look fairly similar too. As usual with sync_io pattern, however, the handler no longer fires from some unknown (to you) thread, but instead only potentially synchronously from within your successful async-wait's call into (*on_active_ev_func)().post() step inside the handler given to .async_request() in the above code..async_request(), there is no need to worry about that possibility: The response always arrives later on an active event that your own code would have detected (e.g., via epoll_wait() or boost.asio .async_wait()). Contrast this with .expect_msg() (et al), discussed below.sync_io patterns, these things will hopefully become second-nature. I.e., none of this is really "special" or particularly applicable to .async_request() or even struc::Channel specifically. We are just trying to be reader-friendly for first-time users.That was quite a lot of stuff; and indeed we've covered at least 50% of the struc::Channel experience, so to speak. That said we've intentionally only demonstrated the sender side of this thing, venturing into in-message analysis only tangentially when it concerned dealing with responses to requests. Now let's jump into the opposing side of these examples, as it is a convenient way to explore expecting notifications and requests in the first place. An analogy: if we'd been discussing how an HTTP GET request works, so far we've covered the client side of it – issuing the GET request among others. Now to discuss the server side – expecting a GET and replying to it.
You might recall from earlier in this page that Flow-IPC imposes a single restriction on the schema Message_body of struc::Channel<Message_body> (in our case Message_body is CoolMsg). That restriction is Message_body must have an anonymous capnp-union at the root. The reason is, specifically, how the .expect_msg() and .expect_msgs() APIs work. Each takes two main arguments:
union-selector (a/k/a union-which value) of the message. Basically it's the root type of the message. E.g., in HTTP the union-which possibilities are GET, POST, etc. In our example they, specifically, are: HELLO_THERE, SUM_REQUEST, SUM_RESPONSE, HIT_ME_WITH_YOUR_BEST_SHOT, MY_BEST_SHOT. (Per capnp convention, these names are generated from the union field names; so helloThere -> HELLO_THERE, and so on.)SUM_RESPONSE and MY_BEST_SHOT-type messages are only used as responses. Therefore we won't .expect_msg*() them. (We could though, if we wanted. See following side-bar.)The union-selector type (generated by capnp) is, for your convenience, aliased as struc::Channel::Msg_which. capnp-familiar people will recognize it as struc::Channel::Msg_body::Which (in our case CoolMsg::Which). It is a C++ scoped-enumeration (enum class) which is light-weight like an int.
nullptr, or not, for the originating_msg_or_null argument to .send(), .async_request(), or .sync_request(). Thus: originating_msg_or_null == nullptr if and only if it's an unsolicited message instance..expect_msg() or .expect_msgs() for that union-which value. It's a response; that's it; either it's expected, or it's not.expect_msg*() is active, then it is immediately emitted to you, the user; if not, then it is cached inside the struc::Channel and may be emitted later once (and if) an expect_msg*() happens to get called by you.expect_*() messages that may also be issued as responses. That is, in the schema, there is no formal way to classify a union-which (message type) as either a response or an unsolicited message; any message can be either. Informally, though, we recommend that you clearly do classify each message as exactly one or the other. (A convention like an Rsp postfix for responses might be a decent idea.) In HTTP, you'd expect an unsolicited GET; but you'd never expect an unsolicited GET-response. We recommend following the same principle in your meta-application and schema. Otherwise it's an avoidable mess..send() or .*_request()); the former by the value of originating_msg_or_null arg to the send-API.That said/understood, the .expect_msg*() APIs are fairly similar to .async_request() – the differences being:
.expect_msgs() one uses .undo_expect_msgs() which is analogous to .undo_expect_responses(). No particular generated ID is necessary: one simply supplies the same union-which value as earlier given to .expect_msgs().We can now demonstrate both .expect_msg*() methods by completing the other side of our one-off SumRequest and indefinite-lifetime HitMeWithYourBestShot examples. As a little bonus this lets us show the use of the originating_msg_or_null arg to .send() (et al). Again – by doing this we classify that message instance as a response rather than unsolicited.
One-off example first:
Indefinite-lifetime variation is essentially similar. In our example the idea is to send responses at various times, not necessarily immediately upon receiving request, so the example code will be somewhat hand-wavy (we are confident you can fill in the blanks).
And that's it! We have pretty much used the entire essential arsenal now. That said a few side-bars:
struc::Channel::Msg_in_ptr which is but a shared_ptr<Msg_in>; while Msg_in is itself an instance of the ipc::transport::struc::Msg_in class template. Its basic capabilities you've already seen: .body_root() to get at the structured data via capnp-generated Reader; ostream << to pretty-print a potentially-truncated un-indented version (potentially slow but okay in TRACE-logging and the like); and so on.Native_handle stored in the original message (struc::Msg_out a/k/a struc::Channel::Msg_out) can of course be obtained from the in-message instance. Use ipc::transport::struc::Msg_in::native_handle_or_null(). One subtlety to note here: unlike Msg_out destructor, Msg_in destructor will not close the Native_handle if any. What to do with this native handle (FD in POSIX parlance) is entirely your call. Do note that in POSIX/Unix/Linux this FD refers to the same resource description as the original sent FD from the origin process; but it is not the same descriptor. The description will itself be released back into the OS's resource pool no earlier than both the original sendable-FD and the received-FD have been closed. So: unless you want that resource to leak, it is your reponsibility to take ownership of it. Hopefully the task is made easier by the sent-message Msg_out (the actual message, not merely the message instance sent) guaranteeing the closure of the sendable-FD no later than Msg_out being destroyed... so the send-side will not be the source of the leak.Msg_in (struc::Msg_in<...>) object represents a message instance. That is it represents a message as sent and received that particular time. A Msg_out can be re-sent later; and it can even be modified and re-sent (as many times as desired). We'll cover that in Structured Message Transport: Messages As Long-lived Data Structures / Advanced Topics – but be aware of it even now. Spoiler alert: If SHM-backing is configured (as we generally recommend), then the original Msg_out and any Msg_in in existence (and there can be multiple) really refer to the same RAM resource. Hence the RAM is returned for other use no sooner than all those guys have been destroyed (across all processes involved). If SHM-backing is not used, then each Msg_in is really a copy of the original, and therefore the Msg_out and all derived Msg_ins are independent, returning each RAM resource on destruction of that object.Native_handles are always, conceptually speaking, copies of the original one(s) (if any) in Msg_out. Again we'll get into all this in Structured Message Transport: Messages As Long-lived Data Structures / Advanced Topics – so consider these paragraphs a brief intro to that stuff.struc::Channel.expect_*()ed, it is cached. A subsequent expect_msg*() may therefore be immediately satisfied by a cached such in-message instance. This is in no way an error. In fact it is an important property: It means no unsolicited message is ever dropped or lost due to not been expected "in time." Also important: any cached messages are emitted to the user – when officially expected and no earlier – in the same order as they were sent. They will not be reordered among each other! The only way to reorder the handling of in-messages is by expecting message type B before message type A, causing the latter to be cached and not emitted in the meantime. E.g., if one expect_msgs() GETs a second after expect_msgs() POSTs, some GETs may be sent before some POSTs but get cached until the somewhat-delayed expect_msgs(GET). However, all those GETs will get emitted to you in a burst and strictly in the order they were sent.expect_*() stuff as early as possible to prevent pathological RAM use..async_request() or .sync_request(), is an error condition but not a channel-hosing one. (A response to .sync_request() received after that call timed out is just dropped; that's it. This is not what we are discussing here.) All such situations – a response to a non-request; a response to a one-off request that has already been satisfied; a response after .undo_expect_responses() – result in the following steps. The response is dropped. If ipc::transport::struc::Channel::set_unexpected_response_handler() is in effect, then that handler is invoked informing you of the unexpected response. Furthermore, via an internal mechanism, the sender-side of the bad response is informed of this situation as well. On that side: If ipc::transport::struc::Channel::set_remote_unexpected_response_handler() is in effect, then that handler is invoked informing you of the unexpected response sent by you..expect_msg*() may trigger a "burst" of cached in-message(s) being emitted via the very handler that was just given to that method; but with sync_io pattern this situation creates a dichotomy: An in-message being available immediately is not quite the same as one being available asynchronously later. Therefore the sync_io-pattern .expect_msg*() API features an extra out-argument. If message(s) is/are available synchronously, it/they is/are synchronously output right into that argument. (.expect_msg(), natually, emits up to 1 in-message, and if 1 was indeed emitted does not register an expectation for more – and forgets the handler, never invoking it. .expect_msgs() can emit multiple in-messages synchronously into a user-supplied sequence container, and even if it does so, it remembers the handler in case more arrive later.)While above we covered all the fun stuff – actual transmission APIs – we glossed over certain less-enjoyable aspects which are nevertheless critical to understand to use struc::Channel. These concern beginning work on a given channel; what errors might occur and how they are emitted; and destroying the struc::Channel properly.
The first thing to grok is that channel operations are divided into 2 categories: incoming-direction ops and outgoing-direction ops. Outgoing-direction ops are simply: .send() and the synchronous work performed by .async_request() and .sync_request(). (.sync_request() is entirely synchronous; .async_request() has a synchronous part as well as the potential response(s) arriving later; only the former part of .async_request() is classified as outgoing-direction.) Incoming-direction ops are everything else and comprise the receipt of unsolicited messages (emission-to-user of which is controlled by expect_*()) and responses (emission-to-user of which is controlled by .async_request()).
Obviously first you need to construct the struc::Channel. We've covered that. (If using sync_io-pattern struc::Channel, you'll need to also call ipc::transport::struc::sync_io::Channel::start_ops() and possibly .replace_event_wait_handles().)
Having done this you can immediately use outgoing-direction op .send() (and .async_end_sending()). You may not use .sync_request() yet; and while you can call .async_request() to send the request itself, no responses will be received or emitted to your handler yet.
You can also call incoming-direction APIs including .expect_*(), but no relevant in-messages will yet be received or emitted to you. To be able to fully use the struc::Channel you must call:
sync_io) .start_and_poll().As this returns, in-traffic may be received and emitted to your handlers (or the special-case blocking op .sync_request() which is unavailable with the sync_io-pattern variant). (These handlers can be registered – after or before the start method – via expect_*(), async_request(), set_*unexpected_response_handler(). Un-register via undo_*() and unset_*().)
The idea of calling a start method is presumably simple enough, but a key input to that method is the on-error handler function you'll need to supply. That guy in and of itself is straightforward enough: it takes a const Error_code&, and Flow-IPC shall set this to some truthy value indicating the error that occurred. However the topic of errors is somewhat subtle. Let's discuss.
One class of error we should get out of the way is the non-fatal a/k/a non-channel-hosing error type. As of this writing there is only one: ipc::transport::error::Code::S_TIMEOUT; it is emitted by .sync_request() if and only if the one expected response does not arrive within the (optional) timeout value given to the method. This is considered a user error, and it does not have any ill consequences as far as Flow-IPC is concerned. (You are of course free to explode everything, if it is appropriate.) If other such error eventualities arise over time, they are documented explicitly. So from this point forward we only discuss fatal a/k/a channel-hosing errors.
The main principle – and one that happens to also apply to various other objects in the library, including ipc::session::Session, ipc::session::Session_server et al, and ipc::transport core-layer senders/receivers – is as follows:
A channel-hosing error occurs at most once per struc::Channel object.
Moreover, for struc::Channel specifically:
A channel-hosing error is emitted to the user at most once per struc::Channel object. Once it has been emitted, it will not be emitted again.
How is it emitted? Answer:
.send(), .async_request(), .sync_request()), then it shall be synchronously emitted by that method call. It is your responsibility to be ready for an error emission.flow::error error emission semantics (themselves inspired by boost.asio albeit a more compact version that reduces the number of method overload signatures by ~50%). The method takes an arg Error_code* err_code.flow::error::Runtime_error exception with .code() returning the specific truthy Error_code being emitted.*err_code to falsy if no error is emitted; else to truthy Error_code being emitted..start() (or sync_io-pattern counterpart, .start_and_poll()). It is your responsibility to take the proper steps inside this on-error handler.This is an important dichotomy to understand. There's at most one occurrence of a channel-hosing error being emitted, and if it does occur, it will occur through exactly one of the above ways. Either a send method will emit is synchronously, or some incoming-direction processing (occurring concurrently in the background with async-I/O pattern, or inside (*on_active_ev_func)() or .start_and_poll() itself with sync_io pattern) will emit it to the on-error handler. Either way you must be ready for it.
What happens to a hosed channel? Rest easy: You can still call its APIs; they will not crash/abort/lead to undefined behavior. They will basically do nothing. In particular most – including .expect_*(), .send(), .async_request() – will return false. One, .sync_request(), will return null pointer (instead of the resulting response message pointer) while setting *err_code to falsy (nominally success) (or not throwing if err_code is null). It is guaranteed, also, that no handler (except that given to .async_end_sending(); see next section) will possibly be called after the channel-hosing error is emitted.
The more salient question is, perhaps, what should you do once you've detected a channel-hosing error (again, either via .send() et al synchronously, or via on-error handler)? Answer: You should stop using the channel; instead you should destroy it (see next section of this page).
You might be thinking to yourself: Well, that's no problem. Channel hosed? Don't use the channel; destroy it sometime soon just to free resources. Indeed. However there is a subtlety albeit applicable only in async-I/O-pattern struc::Channel. It is this:
The channel might get hosed at any time. Consider this code:
Chronologically:
.expect_msg(). Say it returned true which means the channel is not hosed at that exact point in time. It, itself, lacks an err_code arg in its signature, so it cannot itself emit a new error..create_msg() has nothing to do with transmitting anything, and certainyl all the message-reading and -filling logic is orthogonal to the channel itself. No problem..send(). That's a transmission-related API, and moreover it can itself emit a (new) error. Anyway, we don't know if the channel is hosed or not per se, and there's only one way to find out: call it.ok (return value) and err_code (out-arg). (We could have passed-in null or omitted arg – same thing – in which case we'd need to be ready to catch resulting exception.)ok == true, but err_code is truthy. That would indicate new channel-hosing error being emitted. Hence we call teardown() which will stop work on this channel – deinitialize whatever, etc.ok == false. What does it mean? Answer: Something came into the channel in the background concurrently between the start of our in-message handler (when everything was OK) and the point where we decided to send the response. So what should we do? It depends. Informally speaking:ok. Flow-IPC would log sufficiently; trust me.ok == false and return before doing anything else. After all the channel is hosed. In fact this means, with 100% certainty, that the code at Point Y will execute soon. We might as well put all the deinit stuff we want to do, there..send() returned ok == false. Then this code shall ASAP. Presumably something happened on the channel's background in-traffic processing that constitutes the channel being hosed. Usually it's a graceful-close or EPIPE or the like. So we call teardown(), where we centrally handle the deinit of the channel.If one is used to progamming in this async-I/O model – most commonly in our world using boost.asio – this will be familiar. It is a little odd to think of flow control this way at first, but one gets used to it. There are certainly major positives, but this "inverted flow control" could be considered a negative (which people fight in various ways – e.g. by using micro-threads/fibers, though that has its own complexity costs).
That said, if you're using sync_io-pattern struc::Channel, these considerations go away. Nothing happens concurrently in the background, unless your own code makes it so. Yes, there is still the on-error handler; yes, .send() can still emit an error – or return false. However there is no need to worry about the channel being fine at Point A but at Point C .send() returning false. One would "just" not get to Point B, if earlier your own on-async-wait call (*on_active_ev_func)() triggered the on-error handler (synchronously). Flow control becomes linear, and things don't happen suddenly in the background. Hence properly written code should be able to assert(ok) at Point C without fear.
At any rate, with the async-I/O pattern, the above gives a pretty good idea of how to structure error handling and the associated flow control. Basically:
teardown(), called either from on-error handler or upon detecting truthy Error_code from a send-op.false (null + no-error-emission in case of .sync_request()) and possibly short-circuit further logic in that case.struc::ChannelHere we will keep it simple and just give you a recipe to follow. Whether done with the struc::Channel without any error, or due to a channel-hosing error, do the following. These steps could be in teardown() in the preceding example code snippet.
First: Invoke ipc::transport::struc::Channel::async_end_sending(). Give it a completion handler F().
Second: In F(), or after F(), destroy the struc::Channel to free resources. (The Error_code passed to F() is not particularly important – you can/should log it, of course, but that's about it.)
That's it. (In the sync_io-pattern variation of .async_end_sending(), same deal; except that it might (in fact usually will) complete synchronously instead of invoking F(). So handle that synchronously, if it happens.)
struc::Channel? The aforementioned doc header answers that too. Short version: In many cases it's probably fine. However, if the transport out-direction happens to be in would-block state at this time, the opposing side might miss some of the stuff you intended it to receive – hard to say whether that's important or not; depends on your application/protocol's design. So all-in-all it's better to flush outgoing payloads instead of leaving it to chance. This is no different conceptually from properly closing a TCP connection; but truthfully with IPC there's ~zero latency and zero loss, so fewer bad things realistically happen in practice. Still – best to code defensively.sync_io): .async_end_sending() is a Channel-level operation (as opposed to the higher struc::Channel layer). Its completion handler will execute, even if you don't wait for that to occur and destroy struc::Channel first. In that case the destructor will invoke it with an operation-aborted code. If you follow the above recipe, then that won't happen, but defensively-written code should nevertheless do something close to the following in F(): if (err_code == ipc::transport::error::Code::S_OBJECT_SHUTDOWN_ABORTED_COMPLETION_HANDLER) { return; }.The next page is: Structured Message Transport: Messages As Long-lived Data Structures / Advanced Topics.
1.9.4