The changes in #7189 caused an ambiguity between multi-subscribers and
normal subscribers, because the handle type no longer is sent to the network.
Multi-subscribers now go to the network with negative UIDs, normal
subscribers are positive UIDs. UID 0 is never used.
Each client has an incoming queue of ClientMessage.
In the read callback:
- Parse and process only ping messages and a limited number of messages;
anything else will get put into the queue and not processed
- If we queued some messages, we tell the network we stopped reading; this will
result in back-pressure if we are reading too slowly. We also start an idle
handle to process the queued messages.
In the idle handle callback:
- For each client, process just a few pending messages. This is performed in
round-robin fashion across all clients with pending messages
- When a client's queue becomes empty, we re-enable the network read
- When all client queues are empty, we stop the idle handle (so we don't spin)
For local client processing, we use round-robin processing for most cases (including FlushLocal),
but still do batch processing of all local changes for explicit network Flush() calls.
The stream can close (e.g. due to an error) while in the middle of writing. The close callback would immediately destroy the connection object, resulting in the writing code having a use-after-free. Fix this by deferring the deletion to the loop main using a single-shot timer.
Currently in the entire C API of WPILib we have ~8 different ways of handling strings. The C API actually isn't built for pure C callers (We don't actually have any of those). Instead, they're built for interop between languages like LabVIEW and C# which can talk to C API's directly.
For output parameters, the choice was fairly obvious. An output struct containing a const string pointer and a length makes the most sense. Its easy to use these from most other languages, and doesn't require special null termination handling. Freeing these is also easy, as if you ever receive one of these string structures, theres just a single function call to free it.
Input parameters are a bit more complex. To be used from pure C, and from LabVIEW, a null terminated string is the best in most cases. However, null terminated strings in general have a lot of downsides. Additionally, from LabVIEW there are other considerations around encoding that having a wrapper struct helps make a bit easier. From a language like C#, a wrapper struct is by far the easiest, as custom marshalling can make it trivial to marshal both UTF8 and UTF16 strings down.
The final consideration is its nice to have an identical concept for both input and output. It makes the rules fairly easy to understand.
WPILib will not have any APIs that manipulate a string allocated externally. This means WPI_String can be const, as across the boundary it is always const.
If a WPILib API takes a const WPI_String*, WPILib will not manipulate or attempt to free that string, and that string is treated as an input. It is up to the caller to handle that memory, WPILib will never hold onto that memory longer than the call.
If a WPILib API takes a WPI_String*, that string is an output. WPILib will allocate that API with WPI_AllocateString(), fill in the string, and return to the caller. When the caller is done with the string, they must free it with WPI_FreeString().
If an output struct contains a WPI_String member, that member is considered read only, and should not be explicitly freed. The caller should call the free function for that struct.
If an array of WPI_Strings are returned, each individual string is considered read only, and should not be explicitly freed. The free function for that array should be called by the caller.
If an input struct containing a WPI_String, or an input array of WPI_Strings is passed to WPILib, the individual strings will not be manipulated or freed by WPILib, and the caller owns and should free that memory.
Callbacks also follow these rules. The most common is a callback either getting passed a const WPI_String* or a struct containing a WPI_String. In both of these cases, the callback target should consider these strings read only, and not attempt to free them or manipulate them.
We now use a wrapper (wpi::print) to catch exceptions since we can't patch
std::print() to not throw when we ultimately migrate to it.
fmtlib and std format/print throw the same exceptions and always have. We previously patched fmt::print() to not throw a write failure exception, but we can't do that for std::print(); wpi::print() is the migration plan.
The previous fix didn't handle all cases correctly. Instead, add a new
function to raw_ostream (SetNumBytesInBuffer) to allow always using the
full buffer size, and revamp write_impl to more cleanly handle all
cases.
This adds support for two serialization formats for complex data types:
- Protobuf for complex objects with variable length internals that need forward and backward wire compatibility (lower speed, more flexible)
- Raw struct (ByteBuffer-style) for fixed-length objects (higher speed, less flexible)
Deserialization can be done either by creating a new object (for immutable objects) or overwriting the contents of an existing object (for mutable objects).
Implementing classes should provide inner classes that implement the Protobuf or Struct interface (in Java) or specialize the wpi::Protobuf or wpi::Struct struct (in C++). It is possible for classes to implement both. If the class itself does not implement serialization, it's possible for third parties/users to provide an implementation instead.
Uses the Google protobuf implementation for C++ and the QuickBuffers alternative protobuf implementation for Java.
The object was being destroyed due to the error handling path.
Instead, defer to the next loop cycle using a one-shot timer.
Properly handle error return values from Send functions.
Fix UB in accessing one past the end of a vector.
