Arbitrary Hooks

Building a callout API makes sense for common, structured features in software, but occasionally there is a need to provide somewhat arbitrary hook points after the software is designed and the mtev_hooks system is what satisfies this need.

The design goals here are somewhat specific in that we would like to allow for a large number of hook points at low cost when not instrumented. As such, a hash lookup of registered hooks would be considered too expensive. Additionally, we want to provide strong, compile-time type safety as it can be all too easy to hook something with a function using a slightly incorrect prototype that could result in disastrous corruption or crashes (or, perhaps worse, extremely subtle bugs that are punishing to troubleshoot).

The hooks system is simply a set of two macros; one allowing for the declaration of function prototypes for registering and invoking specific programmer-specific instrumentation points, and the other providing an implementation of the registration and invocation routines. Due to the nature of C, the macro calling conventions are less than elegant, but ultimately require no complicated implementation by the programmer.

Hook Declaration

Declaring hooks is done by calling the MTEV_HOOK_PROTO macro with the name of the hook (a term that composes a valid C function name), the arguments it expects, the type of closure (usually a void *), and some variations on those themes that provide CPP enough info to construct an implementation with no programmer "programming."

The declaration of a hook "foo" will result in two functions: foo_hook_invoke and foo_hook_register.

Declaring a hook "foo" (in a header)

This hook "foo" takes a struct timeval * as an argument in addition to its closure.

#include <mtev_hooks.h>

MTEV_HOOK_PROTO(foo, (struct timeval *now),
                void *, closure, (void *closure, struct timeval *now));
Implementing a hook "foo" (in a source file)
#include <mtev_hooks.h>

MTEV_HOOK_IMPL(foo, (struct timeval *now),
               void *, closure, (void *closure, struct timeval *now),
               (closure,now));

Hook Usage

Once the hook is implemented, it can be used by the application and instrumented by code at runtime. In the below example, we'll invoke the foo instrumentation and assuming no issues arise, we'll invoke the original foo_work() function.

Instrumenting a function conditionally

Before we instrument, suppose we have:

  /* preamble code */
  foo_work();
  /* postamble code */

Now we wish to allow programmers to add instrumentation immediately before this code that can conditionally prevent its execution, so we would modify the above code to look like:

  /* preamble code */
  struct timeval now;
  mtev_gettimeofday(&now, NULL);
  if(MTEV_HOOK_CONTINUE == foo_hook_invoke(&now)) {
       foo_work();
  }
  /* postamble code */

If the hook should not conditionally cause or prevent code to run, the _invoke function's return value can be ignored.

In order to register a function that allows the above execution on every other subsequent execution one would provide the following:

  static my_sample_hook(void *closure, struct timeval *now) {
    static int alt = 0;
    return (alt++ % 2) ? MTEV_HOOK_CONTINUE : MTEV_HOOK_DONE;
  }

  void my_init_fuction() {
    foo_hook_register("sample", my_sample_hook, NULL);
  }

The implementation of the hook can be elsewhere in the code, even in a dynamically loaded module. When the hook is registered (you must orchestrate the calling of my_init_function), the behavior of the foo_work() callsite will change and our hook will be called. Given the above implementation, the struct timeval will be ignored, but every other time we reach the call site, foo_work() will be skipped due to a MTEV_HOOK_DONE return value.

results matching ""

    No results matching ""