Dinit Design ============ NOTE: This document is intended to provide an overview. Specifics are generally included as comments/documentation within source files. Design philosophy ----------------- Software design always involves balancing different concerns. Important concerns considered during development of Dinit include: - Usability. The software should be easy to use and understand. Dangerous actions should be guarded (eg stopping a service via dinitctl requires a "--force" argument if it would also cause dependent services to stop). Common use cases should be supported by the feature set. Documentation should be complete. Error messages should be useful. Traceability/debugging is important. - Compatibility and interopability. The software should work together with other pre-existing software; it is a piece of a larger puzzle. New interfaces should only be created when they provide a tangible benefit (eg Dinit supports the same readiness protocol as S6, and works with legacy processes that fork after launching via the bgprocess service type. It logs via the standard syslog mechanism). - Scope. The feature set is restricted, but is large enough to provide utility and usability. - Generality. Arbitrary restrictions should be avoided. Size is kept small and memory use low to make Dinit usable on systems with limited resources. Code remains portable across different operating systems. - Maintainability. Code should be readable and well-documented; it should leverage the compiler's optimisations to remove abstraction penalty, wherever possible, so that the code can be more expressive, and comments should clearly explain why optimisations live "in the code" in cases where it is otherwise necessary. Of course, there are always trade-offs, but I believe that Dinit strikes a good balance with regard to the concerns listed above. Design fundamentals ------------------- The design is based around an event loop: events include process termination, signal reception, and I/O coming in from control sockets (used to issue service start/stop commands etc via dinitctl). The Dasynq library provides the backend event loop functionality; it is included with the Dinit source, but has a separate web page and documentation: http://davmac.org/projects/dasynq/ In Dinit, service actions are often performed by setting "propagation" flags, inserting the service in a queue, and then processing the queues. Thus a lot of service actions are performed by first calling a function on a service and then draining the propagation queues. More information can be found in comments at the beginning of the includes/service.h header. The logging subsystem is designed to log to two different sinks - typically, one is the console and one is the syslog facility, but other arrangements are possible. It uses a circular buffer for each output sink, and if the buffer becomes full it discards additional messages to avoid blocking (though it flags that this has occurred and later outputs an overflow indication to the sink). Services may acquire the console for their exclusive use, and in this case dinit's own logging output is discarded and no longer buffered (though it will of course continue to be logged to the other sink). Control protocol handling uses a circular receive buffer, but allocates storage for outgoing packets dynamically. If allocation fails an "out of memory" flag is set and the connection is closed after writing an "out of memory" information packet. Key considerations ------------------ Dinit's overall function is fairly straightforward but there are some considerations that cause its design to be more complicated than it might otherwise be. The heart of the issue is that Dinit should not stall on I/O and its essential operation should never fail, which means that many code paths cannot perform memory allocation for example, and that log messages all go through a buffer. Note that blocking actions are avoided: non-blocking I/O is used where possible; Dinit must remain responsive at all times. In general operation Dinit methods should avoid throwing exceptions and be declared as 'noexcept', or otherwise be clearly documented as throwing particular exceptions. Errors should always be handled as gracefully as possible and should not prevent Dinit's continued operation. Particular care is needed for dynamic allocations: C++ style allocations (including adding elements to C++ containers, or appending to strings) will raise 'std::bad_alloc' if they cannot allocate memory, and this must be handled appropriately. Functions in 'std::vector' and 'std::string' are also able to raise 'std::length_error' and it should probably be assumed that other containers can do so as well (the C++ language specification is annoyingly unspecific) - although we don't expect to see 'length_error' in practice, it should be handled anyway. Repeating for emphasis: Once it has started regular operation, dinit must not terminate due to an error condition, even if the error is an allocation failure or another exception. Services -------- There are presently five "exposed" service types: internal, triggered, process, bgprocess and scripted. The latter three share the fact that they execute external processes and so naturally share some implementation. The base service class is "service_record" - this directly supports "internal" services (triggered services differ only slightly, but are still implemented via a subclass, "triggered_service"). The "base_process_service" class extends service_record and serves as a base class for the remaining three service types (which all manage external processes) and provides common functionality. Various functions in the service_record / base_process_service classes are virtual, so that they can be overridden by the subclasses. All execution of child processes related to services currently goes through service_record::run_child_proc() (after forking). Services are grouped in a "service_set". This provides the essential interface between the event loop and services. A "placeholder_service" service type also exists; it is used as a placeholder for when a service needs to maintain a link to another service but that other service has not yet been loaded (this is used for example for "after" and "before" service ordering, since those don't require loading the linked service, and "consumer-of" for similar reasons). When any service is loaded, it replaces any placeholder service that already exists for it (and links from other services to the placeholder are updated to refer to the newly loaded service instead). In a similar way, a service which is unloaded but that still has persistent links from other services can be replaced with a placeholder. Source code organisation ------------------------ Most function comments and general documentation can be found in header files rather than in the corresponding source files. dinit-main.cc - just a wrapper around the entry point; reports exceptions dinit.cc - main entry point, command line parsing, general initialisation, event loop processing, signal handling, system management (shutdown / reboot etc) service.cc - base service logic (see service.h) proc-service.cc, baseproc-service.cc - process-based service logic (see service.h, proc-service.h) This builds on functionality in service.cc run-child-proc.cc - contains service_record::run_child_proc(), used to run child processes. load-service.cc - service loading (see load-service.h) control.cc - control protocol handling dinit-log.cc - logging subsystem dinit-env.cc - mainly a function to read an environment file into an "environment" wrapper (as defined in dinit-env.h) tests/ - various tests igr-tests/ - integration tests The utility sources are: dinitctl.cc - the control/status utility dinitcheck.cc - dinitcheck utility dinit-monitor.cc - the dinit-monitor utility shutdown.cc - shutdown/halt/reboot utility Testing ------- The unit tests work by mocking out parts of Dinit, and some system calls, in order to isolate functional units. In the src/tests/test-includes directory are some mock headers. When compiling tests, these headers are put on the include path before the regular includes directory. Note that systems calls are not mocked directly, instead, system calls are wrapped in the bp_sys namespace, as defined in the baseproc-sys.h header; this header is one of the headers mocked for tests. (This avoids problems that might arise from replacing important system calls, and in particular avoids interfering with the test harness itself). It is important when writing new code in Dinit to avoid calling system calls directly, and to instead call the wrapper in bp_sys.