Re: [PATCH v2 0/7] strvec: use size_t to store nr and alloc

On Sun, Sep 12, 2021 at 02:15:48AM +0200, Ævar Arnfjörð Bjarmason wrote:

This is what I'd been sitting on locally since that recent thread, I
polished it up a bit since Jeff King posted his version.

The potential overflow bug I mentioned is in rebase.c. See
5/7. "Potential" because it's not a bug now, but that code
intentionally considers a strvec, and then iterates it from nr-1 to 0,
and if it reaches 0 intentionally counts down one more to -1 to
indicate that it's visited all elements.

We then check that with i >= 0, except of course if it becomes
unsigned that doesn't become -1, but rather it wraps around.

You can also just use ssize_t, or you can compare against SIZE_MAX to
catch the wraparound (there's some prior art in sort_revindex()). That
said, I don't mind rewriting loops to count up rather than down. It
usually makes them easier to follow (and in your patch 5, I do not see
any reason we would need to count down rather than up; we do not even
care where we find "-q", only that we found it.

The rest of this is all changes to have that s/int/size_t/ radiate
outwards, i.e. when we assign that value to a variable somewhere its
now a "size_t" instead of an "int" etc.

I'm a little "meh" on some of these, for a few reasons:

 - anything calling into setup_revisions() eventually is just kicking
   the can anyway. And these are generally not buggy in the first place,
   since they're bounded argv creations.

 - passing a strvec instead of the broken-down pair is a less flexible
   interface. It's one thing if the callee benefits from seeing the
   strvec (say, because they may push more items onto it). But I think
   with strbufs, we have a general guideline that if a function _can_
   take the bare pointer, then it should. (Sorry, I don't have a
   succinct reference to CodingGuidelines or anything like that; I feel
   like this is wisdom we came up with on the list in the early days of
   strbufs).

 - if we are going to pass a strvec, it should almost certainly be
   const, to make it clear how we intend to use it.

So if we we wanted to try to reduce the int/size_t conversions here (and
I don't mind doing it, but am not altogether sure it is a good use of
time, because the rabbit hole runs deep), I think we ought to be
switching to size_t everywhere-ish along whole call chains. Or possibly
providing a checked size_to_int() which will safely catch and abort.
These cases are largely stupid things that real people would never come
across. The real goal is making sure we don't get hit with a memory
safety bug (under-allocation, converting a big size_t to a negative int,
etc).

-Peff