We should be able to honor the fragmentation threshold again. This
passes the maximum bundle size variable to the AAPv2 agent. The value is
determined as the minimum of all maximum bundle sizes reported by the
CLAs and the one specified on the command line, with the special case
of the value 0, which means a maximum bundle size of 2^64 ("unlimited").
Signed-off-by: Felix Walter <felix.walter@d3tn.com>
This enables a clean termination of all agent connections on program
exit without memory leaks that might previously occur due to the "pipe
queue" still holding some data (pointer to allocated memory) when
close() is called on the pipe. For that purpose, the "pipe queue" is
replaced by our standard hal_queue implementation, which also has a
higher performance as it does not require a syscall on every read/write
operation.
As hal_queue does not support poll() (it does not provide a fd), we
cannot check whether the socket has been closed by a subscribing AAP2
client while concurrently waiting for the internal queue. As an
alternative, this implements a liveliness check, which checks the sub.
connection opportunistically from the agent manager: When a new
registration is requested, existing registrations are first
"health-checked" and removed if down, so the new connection can take
over if an old one is down and this was not detected previously.
Based on the new infrastructure, a clean termination routine is
implemented that terminates all tasks transitively by terminating the
listener.
An AAP2 connection handler ("comm task") can terminate in two ways:
1. During runtime, when the connection is closed, it must clean up its own
resources.
2. On program termination, all connections must be cleanly closed and
all outstanding mesages handled.
For 1., we keep using detached threads and allow them to release all the
resources. This commit mainly introduces a mechanism for 2.: A list
tracking all comm. tasks is added that can be used to terminate and wait
for all of them on program termination. A "termination flag" indicates
whether tasks should clean up by themselves or the listener task takes
it over. Everything needs to be protected by mutexes for safe
concurrency.
There is specific handling for macOS, which does not signal other
threads clocking on a socket that the socket has been closed when
calling shutdown() but only when calling close() - thus, close() is
executed earlier on macOS. Also, macOS does not unblock poll() when the
socket is close()d, so we do not use it anymore to wait for data on an
RPC connection, but rather set SO_RCVTIMEO for active-client (RPC)
connections to enforce the keepalive timeout.
Signed-off-by: Felix Walter <felix.walter@d3tn.com>
The backlog configuration option is renamed to indicate up front that it
is intended to be passed to listen().
Signed-off-by: Felix Walter <felix.walter@d3tn.com>
We specify AAP 2.0 to be an RPC protocol whereas the direction of calls
can be switched during configuration. Each call needs to be answered.
This change adds the answer for bundle reception by a client.
Signed-off-by: Felix Walter <felix.walter@d3tn.com>
This is a "meta-agent" based on the current AAPv1 implementation
(`application_agent.c`), which runs the new AAP2 protocol that is
implemented via Protobuf (using NanoPB). Please refer to the
corresponding design documents for further information on the protocol
itself.
Beside the on-wire encoding and future extension capabilities, the primary
difference from AAPv1 to AAP2 is that a connection is now a
unidirectional RPC-like channel. This ensures that acknowledgments
cannot be easily get out of order and makes client implementations
simpler. However, it also means that now a shared secret needs to be
passed by clients to enable registering for the same agent ID.
The following limitations of this first implementation are present:
- Only sending and receiving ADUs is supported.
- The shared secret is not checked - any app can register for the same
agent ID as long as no other app has yet registered for the requested
direction (subscribe or non-subscribe).
- Only one client per direction is supported at the moment. (It is at
least planned to support multiplke subscribers for an agent ID.)
- The server does not expect acknowledgments from clients when they
receive ADUs.
- The keepalive mechanism is not implemented.
In general, this reflects the feature set of AAPv1, but with massive
extension opportunities.
Signed-off-by: Felix Walter <felix.walter@d3tn.com>
This adds the Nanopb library as submodule under `external/` and adds
corresponding build configuration to the Makefiles. Additionally, a
Protobuf stub for AAP 2.0 is included along with a `make` target to
build the corresponding C language files (`aap2-proto-headers`).
Signed-off-by: Felix Walter <felix.walter@d3tn.com>
