Hacker News Reader

Comment by adrianN | Hacker News Reader

Imo every piece of code (for reasonable definitions of "piece") is supposed to check its own preconditions and not rely on the caller to check them.

View on HN

It's a bad idea, especially in kernel/low-level/performant code. It's OK to check some values at specific points (like when a value is passed from user space), but in general it's bad practice to check it at every single function. You trust in your program flow.

Imagine if at every function down a complex stack you go with:

    if (!ptr1 || !*ptr1 || *ptr1 > 5 || param1 < 0 || param2)
        /* etc.... */
    {
        return NULL;
    }

(used arbitrary names and values).

This is nice in theory but a bad idea in practice (as a blanket statement, I am all for checking preconditions in general). An easy example is binary search, which I think is a reasonable "piece" of code by your definition. One should never check its preconditions _inside_ the binary search function (that the list of elements being searched is partitioned by the search predicate). Checking the precondition is O(n) while the algorithm is O(lg n).

You're right, but checking for null is not terribly expensive compared to iterating over a linked list.

The thing is that you aren’t really advocating for just that... Since this is pretty general support library code, the result of this philosophy is defensive checks in all support functions and that means checks every spinlock, mutex/semaphore, and god knows how many other places... everything really. We also would want these checks at multiple levels since we shouldn’t trust that other people checked things. This would have a significant effect on performance and many would prefer their system not to run however-many percent slower. All to avoid a possible issue that can’t be very large or systems would be kernel panic’ing all over the place.

From a black box, zero-knowledge, perspective it’s maybe worth remembering that code built this way successfully runs mission critical systems all over the world every day. Thoughtful people would have switched to other (slower, more pedantic) systems if doing things differently was a real overall life improvement. People have had the choice, there are plenty of OS’s out there... some far more formal/pedantic. Linux wasn’t the safe choice back in the day, it was just better by several important metrics consistently over time.

Perhaps these insanely experienced kernel devs know what they are doing.

This isn't possible in many cases - consider the simple C library function strlen, one of who's preconditions is that it must only be called on a zero-terminated string. There is no way to write code in strlen to check that this is true.

Which is one of the reasons why coding standards like MISRA forbid using strlen (at least in C++, I guess even MISRA doesn't want to force you to write safer C with sane string types).

Reply on news.ycombinator.com