totally_safe_transmute, Line-by-Line (2021)(blog.yossarian.net) |
totally_safe_transmute, Line-by-Line (2021)(blog.yossarian.net) |
Note the emphasis on "default" above; you can use the Linux sandboxing features such as seccomp-bpf to build a sandbox which is truly memory-safe, closing this hole. The OS is in charge of the features it exposes, and Rust can't do much about that.
Note also that the existence of totally_safe_transmute doesn't mean that Rust's memory safety features are pointless. Empirically, memory-safe programming languages result in far fewer memory safety vulnerabilities, because they make exploitation way harder.
And even then, it doesn't help with whole-system security when every program in every other language on your system, including safe ones like Java and Python, have the same capability. (Although if I'm wrong that there's no precedence for languages attempting to block this, I'd love to see the prior art.)
That's the easy part, you'd provide some OS-specific API for getting this file handle, call it `std::os::fs::proc()` or thereabouts, and make it an `unsafe` function. The "hard to do" is the bigger problem, because if you're providing an unsafe alternative then you'd like to plug all the holes in the safe interface, which AFAIK is non-trivial.
There’s an updated version with Windows support and better performance: https://github.com/John2143/totally-speedy-transmute/
What worries me is this macro, which “smuggles” the unsafe keyword past the forbid(unsafe_code) flag: https://github.com/John2143/totally-speedy-transmute/blob/ma...
In my mind, this kind of capability makes Rust crate safety scanning and associated metadata worthless as currently implemented.
Package management tools ought to store code instead of binaries, and perform safety checks to via instrumented compilers.
Every day that goes by is a day I think we should make a beeline to CHERI even when we have "safe" languages.
Casts are conversion: a new value is produced based on an existing one.
Reinterpretation requires a value to be in memory, and to be accessed using an lvalue of a different type. Most situations of this kind are undefined behavior.
double transmute_int_to_double(int x) {
double result = 0;
memcpy(&result, &x, sizeof(int) < sizeof(double) ? sizeof(int) : sizeof(double));
return result;
}
transmute_int_to_double:
movd xmm0, edi
ret
Not by my interpretation of n3096 (April 2023 ISO C draft).
> doesn't even necessarily touch any memory
The abstract semantics calls for memory being touched. Data flows that go through memory in the abstract semantics can be optimized not to go through memory. UB can do anything at all.
If you wanted to backdoor a Rust program, you wouldn't need the `unsafe` keyword at all. And if you want to use unsafe code, that's fine, plenty of crates use unsafe code without anyone being up in arms about it (e.g. the regex crate). This is a party trick rather than something to be concerned about; at the end of the day either you're auditing your dependencies (in which case this would stick out like a sore thumb) or you're not (in which case there are far easier ways to pwn you).
The totally safe transmute is at runtime, so an instrumented compiler cannot detect it (halting problem is in the way). You'd need runtime instrumentation of your binary. And even then, it's wildly impractical.
If you let an application interact with the environment, tomorrow Linux or Windows could add a new magic file, or a special COM call, or whatever it is that creates unsafe. Rust can't have a complete list of all the unsafe things that are outside of its control.
What you probably want is a runtime VM, like WASM.
The linked updated library uses a different method: it literally smuggles the "unsafe" keyword past the safety checks by removing the space character from "un safe".
This can and should be caught by the compiler -- it has full access the syntax tree at every intermediate stage of compilation! Instead, the Cargo tool and the rustc compiler are simply keyword-searching for the literal string "unsafe", and are hence easily fooled.
Note that this updated method is not the same thing as the Linux process memory mapping and doesn't rely on OS APIs in any way. It is a purely compile-time hack, not a runtime one.
What I'd love to see is an healthy ecosystem of open-source packages that are truly safe: using a safe language, without any escape hatches, etc...
E.g.: Libraries for decoding new image formats like JPEG XL ought to be 100% safe, but they're not! This hampers adoption, because browsers won't import huge chunks of potentially dangerous code.
Now we can't have HDR photos on the web because of this specific issue: The Chromium team removed libjxl from the code base because of legitimate security concerns. A million other things are also going to be "unsupported" for a long time because of the perceived (and real!) dangers of importing third-party dependencies.
We'll be stuck with JPG, GIF, and PNG forever because there is no easy way to share code that is truly safe in the "pure function with no side-effects" sense.
PS: This type of issues is also the root-cause of issues like Log4J and various XML decoder security problems. By default, most languages allow arbitrary code even for libraries that ought to perform "pure transformations" of one data format to another, such as reaching out to LDAP servers over TCP/IP sockets, as in the case with Log4J. Similarly, XML decoders by default use HTTP to retrieve referenced files, which is madness. Even "modern" formats like YAML make this mistake, with external file references on by default!
> What you probably want is a runtime VM, like WASM.
Sure, that's one way to sandbox applications, but in principle it ought to be entirely possible to have a totally safe ahead-of-time compiled language and/or standard library.
Rust is really close to this goal, but falls just short because of tricks like this macro backdoor. (It would also need a safe subset of the standard library.)
The LLM hype has really jumped the shark.
ETXTBSY.
How about ptracing yourself?
How about undervolting your CPU such that it malfunctions?
How about running your program off a FUSE filesystem that changes the contents out from under it?
How about editing the raw block device you’re running from?
How about modifying /dev/mem on older systems that allow that?
This particular rabbit hole is extremely deep, and I don’t think it’s practical to address it for real short of making a real security model for what running code may and may not to do the rest of the world. (I think doing that is an extremely worthwhile endeavor, but disallowing opening specific files won’t be part of it.)
> This can and should be caught by the compiler -- it has full access the syntax tree at every intermediate stage of compilation! Instead, the Cargo tool and the rustc compiler are simply keyword-searching for the literal string "unsafe", and are hence easily fooled.
This is not what's going on. No Rust tool performs any such string scanning. The "un safe" with the space is purely for aesthetic effect. In fact, the proc macro works exactly the same when modified to just use "unsafe" directly, and cargo-geiger still doesn't report any issues.
The real effect is that macros and proc macros from foreign crates are allowed to emit unsafe { ... } blocks in their output, even if the caller of the macro uses #![forbid(unsafe_code)]. Effectively, the output is considered as coming from "different crate" from the caller with the #![forbid], so it's not effected by the forbidden lint. This behavior was implemented in [0], which silences all lints (including forbidden lints) within the output of macros defined in foreign crates, except for certain lints which explicitly opt-in.
The #![forbid(unsafe_code)] inside the src/lib.rs defining the proc macro similarly doesn't do anything, since it restricts the definition of the proc macro, but not the output of the proc macro.
As far as I know, there is currently no way to deny unsafe code in the output of proc macros.
What clause(s) support the claim that that example is UB? At least at first glance it looks pretty similar to the sometimes-recommended way to perform safe type-punning in C, and the only way to "directly" invoke UB via memcpy (passing pointers to overlapping objects) isn't relevant here.
The only suspicious part to me is the size/number of bytes to copy, but I'm not sure that's outright UB?
That being said, is there a representation for doubles that could be considered invalid? I thought all double bit patterns are interpretable as a "valid" double, even if you get NaNs.