Also, it's not just about vulns in security-critical code. It's also about ordinary bugs. Why not be a little more careful? It won't hurt.
> It might get unzeroed if you work with the memory.
Maybe, but it isn't very common. I'm not sure when the C standard allows changing padding bytes, but in practice the compilers I've used don't seem to do it. And again, it's just a debugging aid, if it causes too much trouble on some platform, just turn it off.
If the extra code is really that bothersome, just use a macro or wrapper function.
$ clang -x c -
#include <stdlib.h>
int main(int argc, char **argv) {
char *foo = malloc(10);
free(foo);
free(foo);
}
$ ./a.out
a.out(14391,0x1024dfd40) malloc: *** error for object 0x11fe06a30: pointer being freed was not allocated
a.out(14391,0x1024dfd40) malloc: *** set a breakpoint in malloc_error_break to debug
Abort trap
As you turn up your (automatic) checking this will be caught more and more often. Setting the pointer to NULL will silently hide the error as free(NULL) is a no-op and nothing will catch it. Thus, the suggestion here was1. advocating adding additional code, which has historically been hard to actually do in practice, and
2. providing a suggestion that is generally worse.
I can see an argument for wrapping it in a macro so you can turn off nulling in debug builds (ASan might even have hooks so you can automate this, I know Valgrind does). But use-after-free is worse than just double-frees, and if you read a dangling pointer in production there's no real way to catch it AFAIK. Last I heard (admittedly been a few years since I checked), you're not supposed to deploy ASan builds because they actually increase the attack surface.
So, your program's memory is full of these dangling pointers, and at some point you will have a bug you didn't catch and use one. And you can't even write an assertion to check that it's valid. What do you propose?
And again to clarify, I'm not trying to advocate for hiding bugs. I want to catch them early (e.g. with assertions), but I also want to avoid reading garbage at runtime at all costs, because that's how programs get pwn'd.
From an exploitability point of view they are largely equivalent.
As for the rest of your comment: my point of view is largely "you should catch these with Address Sanitizer in debug", so I don't usually write code like "I should assert if I marked this as freed by NULLing it out". If I actually need to check this for program logic, then of course I'll add something like this.
The macro you suggest would alleviate my concerns, I suppose, and it wouldn't really be fair for me to shoot that solution down solely because I personally don't like these kinds of assertions in production. So it's not a bad option by any means, other than my top-level comment of this requiring extra code. I know some libraries like to take a pointer-to-a-pointer so they can NULL it out for you, so that is an option for your wrapper. And a double-free that doesn't crash can sometimes open up exploitable bugs too since it messes with program invariants that you didn't expect. But these are much rarer than the typical "attacker controlled uninitialized memory ended up where it shouldn't" so it's not a big deal.
That's not what the parent comment said.
You retort: "I have yet to see code that has remained secure solely on the “discipline” of programmers remembering to check things."
I think that is a dishonest misrepresentation of what the parent comment said, isn't it?
You took the original argument, changed it to be fallacious, and used it as a strawman. That's what was dishonest about it.
Left unsaid due to the fact I was out doing groceries this morning when I posted that was that I don't think this is even a very good practice in general, as I explained in more detail in other comments here.
Honestly, unless you really cannot afford it time-budget wise, I would just ship everything with ASAN, UBSAN, etc. and deal with the crash reports.
This is like the most clichéd way of saying “my code has security vulnerabilities” that there is. I have yet to see code that has remained secure solely on the “discipline” of programmers remembering to check things.
> The trick of checking that buffers are zeroed is purely a debugging tool, so it's okay if it doesn't work on some platforms.
Fair.
> And if you allocate with calloc(), the padding will be zeroed for you.
It might get unzeroed if you work with the memory.