Jeremy Cole <jeremycole@google.com> writes:

The attached patch adds what amounts to compile-time checking for unmatched DBUG_ENTER/DBUG_LEAVE (DBUG_RETURN, DBUG_VOID_RETURN) by introducing a variable in DBUG_ENTER which is only used in DBUG_LEAVE. This allows any compiler which can robustly detect unused variables to detect the mismatch at compile-time. There is already a run-time check for this case, but it is somewhat limited as it requires _db_return_ to be called in order to detect the mismatch, and this is in practice not always the case. Particularly three cases allow this to escape detection:

Ideally we'd be able to check for gcc attribute noreturn but I'm not sure there's a way to do that (as it's not a preproceessor thing). I'm pretty sure that with some optimization levels that the temp variable can get optimized out (although I haven't looked at the produced code). Personally, looks good and I'm tempted to pull it into Percona Server. -- Stewart Smith

Hi, Jeremy! On Jul 19, Jeremy Cole wrote:

Prior to posing it as a final patch contribution or opening a JIRA ticket, I wanted to propose this idea.

The attached patch adds what amounts to compile-time checking for unmatched DBUG_ENTER/DBUG_LEAVE (DBUG_RETURN, DBUG_VOID_RETURN) by introducing a variable in DBUG_ENTER which is only used in DBUG_LEAVE. This allows any compiler which can robustly detect unused variables to detect the mismatch at compile-time. There is already a run-time check for this case, but it is somewhat limited as it requires _db_return_ to be called in order to detect the mismatch, and this is in practice not always the case. Particularly three cases allow this to escape detection:

1. Some instrumented functions are called only from non-instrumented functions. 2. Some instrumented functions are called only from non-returning functions, such as abort or die functions. 3. Some instrumented functions are called from functions who call my_end or my_thread_end (making cs unavailable) before returning.

In any case please see the attached DRAFT patch for a full implementation of the idea. Note one bit of ugliness is that non-returning functions, after this patch, must be marked explicitly as DBUG_ENTER_NO_RETURN -- this actually is not really a bad thing, though in and of itself.

Thanks, Jeremy! I kind of liked the idea. Although I, probably, wouldn't use DBUG_ENTER_NO_RETURN, but a separate macro DBUG_NO_RETURN that would use your guard variable without calling _db_return_. Hmm, come to think of it, perhaps functions that don't return should still use DBUG_LEAVE? Then no special new macro is needed at all. But what made me thinking was, that this patch only detects whether the function has at least one DBUG_RETURN, anywhere in the function body (your changeset comment mentions this, albeit in a wrong place, so it's a bit confusing). In my experience a notably more popular mistake is to miss DBUG_RETURN only somewhere in a function, like in the early return on the error. What could we do to solve this too? Is C++ class the best we could have? Regards, Sergei