Compiler Can Undo Your Security Checks
davidbombal.com
davidbombal.com
https://cppreference.com/c/string/byte/memset (memset_s)
https://cppreference.com/cpp/string/byte/memset (memset_explicit)
Also see notes section for various other functions that people were using before memset_s became standard.
Security properties on the other hand are very often about the relationship between these sequences. For example we like to say that an external observer/attacker can’t distinguish internal state by the external observations (confidentiality) which requires that two observable traces given different hidden values don’t have different observations from the same starting observable state.
If you’re an LTS nerd you know this difference as trace properties vs hyper properties. Compilers try to preserve the former but not the latter.
Separately there is the problem of “what is observable?” For example, if you include timing in your observable behaviors suddenly the kinds of compiler passes that are able to preserve observations tends to zero rather quickly.
Whoa. That's not even something it should be able to detect. What's the gaping security hole that lets it do that?
the most common way is to set a breakpoint, which pause JS execution only if devtools are open. this pause can be detected with a timer.
Well, it sounds like a lot of compilers are making unjustified assumptions about what the outside world is allowed to affect or observe. Maybe with the encouragement of specs, maybe not.
The compile can transform your code if the result of the computations it makes is the same (plus any ordering guarantees you've encoded with the correct primitives, etc.). "Interact with the outside world in exactly the same way" is way too strong a guarantee.
The canonical example is the constant-time comparison. You want to compare a provided password hash with the one in your database in such a way that every comparison, regardless of success or failure, completes in exactly the same amount of time. The compiler does not care about this desire of yours, though, and can and will try to optimize things so it will stop the comparison as soon as it knows they don't match, which can take different amounts of time depending on the input. This is perfectly valid and reasonable, but breaks security sometimes.
Another example is to allocate some memory, write to it, and free it, without reading it. The compiler is free to optimize the entire thing away. We have `volatile` in C because sometimes merely writing to a memory address has side-effects that aren't visible to the compiler, but if you don't use it, the compiler can do what it wants.
The word "result" is doing a lot of work there. `printf ("%d\n", 2+2);` isn't interesting because 4 appears in a memory cell; it's interesting because 4 appears on stdout. Which one is the "result"?
If you're going to make assumptions about what's "inside" the program and what's "outside", you have to make them explicit. And they have to be reasonable assumptions. An assumption that memory is "inside" has to be justified in the presence of shared memory, virtual memory, debuggers, or whatever. You have to actually explain what you mean in a lot more detail than I think the average spec has a chance of doing.
If I create an unlinked temp file, and the compiler can observe that I'm holding the only FD open on that file, should that file be seen as "inside the computation", or as "a collection of results and inputs"?
Without reading the specs, I can be 95 percent sure that they don't nail down all the issues... and 100 percent sure that if they do nail down all the issues, or even all the possibly important issues, the corner cases are unknown to almost all actual programmers. Which means either that it's not appropriate for the compiler to rely on just any rule regardless of what the spec says, or that it's not reasonable to write code in the language.
Well, in C/C++, as soon as your program has one UB bug, the compiler has absolutely no obligation whatsoever.
Rust also has UB, btw, https://doc.rust-lang.org/reference/behavior-considered-unde..., so I don't know where it is that people have imagined this is something the C and C++ language designers went out of their way to foist upon you.
If you want to write code for a VAX, then you can use the K&R C compiler where it had defined outcomes for everything. If you want to write portable C code for modern CPUs then it's fair to ask what the C language standard is supposed to define for each of those CPUs and OSes and ABIs.
And a bunch of people were nice enough to do that for you and I, but because they are not deities, there are things that they had to leave out to make the language useful, so they did.
Not at all the same. C and C++ are full of hazards and I get the impression it’s genuinely difficult to avoid entirely in normal code bases, and typically impossible to avoid statically. Whereas in Rust it’s all gated behind the unsafe keyword, and if you don’t use it (and most code bases never need to use it), you cannot encounter UB; and that scoping makes it far easier to control and handle correctly.
Use-after-free and race conditions are the areas where Rust has a clear advantage. Here one needs to have a clear strategy and enforce it manually (or using tools, but we lack good open-source tools for this). Valgrind and similar tools also help.
Rust protects against all of these, but if you think Rust is only about memory safety, I don't believe you have seriously tried it. It does a lot of things in std API design as well to steer you away from bugs. Some examples:
- The pervasive use of Result and Option makes it impossible to forget to handle (or forward) the error case.
- Because of usage of RAII (C++ has this too, but not as pervasively, C doesn't except using some very new GCC extension) it is very hard to forget to free resources such as files, sockets, database connections, mutexes, etc.
- Enums can carry payload in their variants (C devs: think tagged unions, but safe, C++ devs: think std::variant but with match/case rather than bulky visitor pattern), which means you can make API designs that cannot represent invalid states.
- The typestate pattern is a bit hard to explain briefly, but it allows a state machine with types at compile time, to make sure you dont misuse an API. For example it can be used to prevent forgetting setting required fields in a builder before building. Or in embedded microcontrollers to make sure you can't hand out the same GPIO pin to different parts of the code base.
I often find that my code in Rust works first try, while that almost never happen in C++ for non-trivial code. It is what all those Haskell devs were talking about all these years, but in a systems language (no GC is critical to my day job in hard realtime control systems) and without the incomprehensible abstract math lingo.
You can do a lot more in C too: You can design safe interfaces based around incomplete structure types. This also should allows what you call typestate pattern (if I understand it correctly). You can build a decent option type / result type. You can have a bounds safe vector type. You can have safe string types. One can have type-safe dynamic casts. One can annotate return values so that they can't be ignored. One can use many different tools for safety. People coming from C++ often think that one can not do this in C because "it lacks abstractions", but this is not really true.
EDIT: Also, errno is an awful design. Forgetting to check for errors, or screw up which error you report is so easy in C. Exceptions in C++ are also bad, it is very easy to have no idea what exceptions are possible 5 layers down and end up with unhandled exceptions.
In Rust I get good defaults, and a language that guides me in the right direction. The rules for safety critical rules are somewhat still under development but so far they look a lot shorter (you still need the "don't allocate in hard realtime tasks except at startup" and similar rules for example). And if you aren't doing safety critical you can safely use all of the language as long as you stay away from unsafe.
And for most code you dont need unsafe, and even when you do someone else has likely done the hard work for you already, providing safe abstractions on top. (The exception is FFI to other languages, it is impossible to avoid unsafe when calling code in another language that the compiler can't reason about, you should build a safe Rust API on top of the raw bindings. For popular libraries this often already exists.)
That you do not get safe libraries out-of-the-box in C is a major problem. But I also see the supply chain situation in the Rust world as highly problematic.
There are also experiments in formalizing the safety comments with attributes. To me the current prototypes look halfway towards formal verification, with unsafe functions specifying named requirements that must be upheld when calling them and the callsites needing to "discharge" them by name. Time will tell if this is a good idea for general code and it it catches on.
If even they had to add escape hatches despite the presences of powerful language primitives like types, traits, borrow-checking, maybe the people charged with making it all work with 80s compiler technology weren't the literal Antichrist for also having UB as Rust does.
For what it's worth, C and C++ are much different in terms of hazard, so when you bucket them together it makes me wonder how familiar you are with the actual risk of UB in practice.
Nowadays it generally stems from doing weird things.... but no one is making you do weird things, any more than people are making you use Rust's unsafe keyword.
The extent to which this is about culture should not be underestimated, to me that's the most hilarious part of Bjarne Stroustrup's big rant on memory safety a few years ago. The C word appears exactly once in his slides, in a quote from somebody else about what needs fixing. But Bjarne never addresses this once, even though it's the actual problem.
They did. Most other languages, the vast majority of which are also memory safe languages, go out of their way to do the opposite, and give meaning even to erroneous programs. Some things slip through the cracks, and generally language designers and implementers work hard to get rid of UB.
C/C++ is the only ecosystem that has fully embraced UB as a way of life. They are the only compilers that make full use of "UB is bad and cannot ever happen" as a core tenet in how optimizations are designed. Rust UB is at least a little different. Rust UB can only be the result of unsafe code and is meant to be limited in blast radius, and is absolutely not meant as a loophole for compilers to just do whatever to make the code faster.
C/C++ have a surprisingly large set of UB, too. Thankfully, the rest of the software world is rising up and the committees are starting to make things like gasp signed arithmetic overflow into defined behavior.
But don't hold your breath.
But note that UB also does not necessarily mean your program has no meaning. A good compiler can do something reasonable by defining the UB. There is no mandate in the specification that a compiler has to break things. It is also the user's responsibility to put pressure on compiler developers to do something reasonable.
However, how many years it will take until those versions become widely deployed across major compilers, and used by developers?
This isn't true. Rust UB is meant to only be possible to trigger via `unsafe` code, but if you do trigger it, the compiler is free to do whatever it wants, and in practice it will make full use of that freedom. Rustc uses LLVM, it shares most of its optimizations with Clang!
We can even rope off whole parts of the software. If the Postscript printing code has UB, simply instruction operators only to use the HP inkjet printers for which we know Postscript is not used can prevent this UB from happening.
undefined behavior: "behavior, upon use of a nonportable or erroneous program construct or of erroneous data, for which this document imposes no requirements" The C++ spec at some point changed this to explicitly allow changing anything in the program not just the specific behavior implied by the "for which". The C spec never did this.
Because people were confused about this, in C23 we added the following note. "Note 3 to entry: Any other behavior during execution of a program is only affected as a direct consequence of the concrete behavior that occurs when encountering the erroneous or non-portable program construct or data. In particular, all observable behavior (5.1.2.4) appears as specified in this document when it happens before an operation with undefined behavior in the execution of the program."
Compilers mostly follow this. GCC has bugs related to volatile. Clang often follows the C++ standard, where it is different from C, so probably does not conform to the standard here (as for some other things).
The new C++ standard will have UB "barrier", i.e. std::observable that will limit the effect of UB to things before this barrier.
> Volatile operations are permitted to trap. The compiler may not assume that execution will continue after a volatile operation.
To some extent in C and even more in C++ there's a much worse problem, IFNDR [Ill-formed No Diagnostic Required]. Programs which the language specification insists mean nothing at all, but your tools won't (in many cases can't) notice so the result might do anything. It's not Undefined Behaviour, your program never had any defined behaviour at all.
It’s worth being clear here that this is not what Fil-C does, it still has UB, and can still explode in many of the same ways as C and all (after all, it’s a clang fork). Fil-C takes one particular class of allocation related bugs and UB off the table, but leaves many of them behind.
Leaks could reasonably be considered unsafe as they can cause a program to crash due to resource exhausting, even where the actually used memory is limited.
If you do not consider leaks as part of the problem, there is a trivial way to avoid use-after-free and double-free: Simply never free any memory or only at the very end. (Which in some scenarios is exactly what people do.) The later would fulfill your definition of "holding references to be able to free them".
You can make this claim about any resource, there's no reason to single out memory here. You can run out of file descriptors, inodes, connections to a remote database, disk space, anything - including CPU time.
And notice that it wasn't the leak that you've now said was unsafe, it was the use itself. The program didn't blow up "because of a leak", it blew up because we exceeded some arbitrary resource threshold which may have been invisible to us.
> If you do not consider leaks as part of the problem, there is a trivial way to avoid use-after-free and double-free: Simply never free any memory
Indeed. And that's exactly what we see in some domains and if you've solved the other issues (e.g. bounds misses, type confusion) you've now got memory safe programs. Most general purpose software can't be written this way, but there is a whole heck of a lot of software out there which could be.
> The later would fulfill your definition of "holding references to be able to free them".
And that former would be characterized as "leak everything" and indeed that's entirely safe and, just as I said, to an end user this is a distinction which makes absolutely no difference.
To the end user memory leaks are not important as long as you do not exceed the available memory and the program does not crash. The moment it does, it a problem and it can be a risk. What I agree with, (if you argued that point - but you don't), is that it is more benign and qualitatively different risk than unbounded behavior you may get with an out-of-bounds access or use-after-free.
Are you claiming that there must be such transforms or only that it is likely that bugs exist which isn't interesting.
To assert that something specific always happens under UB is counter to the definition of UB. Fil-C carefully defines away UB for some operations, but to make full guarantee of safety, even by their definition and modulo bugs, I believe it necessary to fully remove UB.
Fil-C gets to look at some code which has UB if variable z is 9 and go "OK, lets check whether z is 9, and if so...". The result, of course, is much, much slower than executables from a typical modern C compiler, but since "It always does X" is in fact a permitted implementation of "Undefined Behaviour" this is a compliant C implementation, at least in most observable respects.
#include <stdio.h>
#include <stdlib.h>
int shift(int x, int n) { return x << n; }
int main(int argc, char **argv) {
printf("%d\n", shift(1, 32)); /* n == width: UB */
return 0;
}
static void loop(void) {
int s = shift(1, 32);
int n = 0;
for (int i = 0; i < s + 3; i++)
n++;
printf("[loop] iterations=%d (s+3=%d)\n", n, s + 3);
}
static void sw(void) {
switch (shift(1, 32)) {
case 0: puts("[switch] case 0"); break;
case 1: puts("[switch] case 1"); break;
default: puts("[switch] default"); break;
}
}
int main(int argc, char **argv) {
loop();
sw();
return 0;
}
You said "explode" earlier and so I was expecting something a bit more dramatic than "Unsurprisingly Fil-C has unspecified results for some expressions".
The behavior of the program itself when undefined behavior is invoked is allowed to be _anything_. I've simply demonstrated here that the compiler is using the fact that there is undefined behavior to perform optimizations that would not be allowed without undefined behavior. Those optimizations are allowed to result in the program doing anything at the compiler's whim.
I spent a few minutes poking just to see if my gut is right here, and already, here's an example of UB being used in an optimization by the compiler that leaves a fil safety check at on -O0 but drops it at higher optimization levels. I find it difficult to believe that all of the complex interactions of every optimization pass in the presence of even this subset of UB are guaranteed not to violate these memory safety promises.
#include <stdio.h>
#include <stdlib.h>
__attribute__((noinline)) static void poke(int *p, int k)
{
int n = 32 + (k & 15); /* always >= 32: shifting an int by >= 32 is UB */
p[(1 << n) * 20] = 0x41414141; /* on x86 the CPU computes index 40, out of bounds */
}
int main(int argc, char **argv)
{
int *p = calloc(16, sizeof(int));
poke(p, argc);
puts("after poke");
return 0;
}I also do not believe that all complex interactions are guaranteed to not violate all safety promises. I also know that this is not true for Rust, so what is your point?
Here, the issue seems that compilers can reload variables. If this is a bug, then you already have a data race in your program which you can prevent with correct use of locks and/or atomics.