The last thing the world needs is another immature SSL/TLS implementation however this makes it very interesting:
> No dynamic allocation whatsoever. There is not a single malloc() call in all the library. In fact, the whole of BearSSL requires only memcpy(), memmove(), memcmp() and strlen() from the underlying C library. This makes it utterly portable even in the most special, OS-less situations. (On “big” systems, BearSSL will automatically use a couple more system calls to access the OS-provided clock and random number generator.)
> On big desktop and server OS, this feature still offers an interesting characteristic: immunity to memory leaks and memory-based DoS attacks. Outsiders cannot make BearSSL allocate megabytes of RAM since BearSSL does not actually know how to allocate RAM at all.
> The last thing the world needs is another immature SSL/TLS implementation...
No, they world needs as many of those as it can get (if they are tagged as such). And then all the authors need to discuss with each other what they learned from the implementation. And then they compare their code with existing code bases and discuss differences. They review their patches, as well as patches from other projects.
They discuss the diffs, learn from others and bring in new ways too look at things.
All bugs are shallow given enough eyes, and yet one of our biggest mantra's sole purpose is to limit the number of eyes. We need people that are familiar with crypto codebases and its subtleties, because we need the reviewers for our established projects. And for this reason, we need people to write and publish crypto related code. Not so we can push another new, excitingly half-baked TLS stacks into a product, but to foster the code review process we all rely on.
I'm not sure that was the issue with OpenSSL. According to the libressl folks the OpenSSL team were spending massive amounts of time on FIPS support at the expense of known serious issues the OpenBSD team had raised.
What sorts of applications are written for OS-less systems that require a TLS library?
EDIT: Thanks for the sincere responses. In retrospect my question might have appeared smarmy, but that wasn't my intent and I really appreciate the responses.
You don't need TLS for that, you could simply use an HMAC and a shared secret, assuming you're not worried about people with physical access (and ability to get the secret) being able to create updates. Of course if you've got multiple instances of the device (not some hobby thing where they are all owned by you), then the secret for each device should be different so someone can't buy the device, determine the secret and then push updates to other people's devices.
RSA means big integer which means unhappy performance on devices that often don't even have floating point in hardware. I think elliptic curve could be faster?
That question doesn't make any sense. They don't do the same thing, so you can't compare their relative performance. What's faster: an Intel i7 or a BMW i8?
Malloc is potentially troublesome for two reasons. First, its performance is potentially unpredictable. It depends on the current state of the heap at the time of the call, which you can't know in advance except in some very rare situations. It can also fail entirely, and that is likewise unpredictable.
memcpy and memmove, ultimately being byte-copying loops, don't suffer from these problems. Their performance is consistent and they always succeed if your pointers and lengths are valid.
On PCs these days, the troubles of malloc don't matter much. You have so much performance margin that occasional slow calls don't matter, and virtual memory with a big address space means that it almost never fails. If it does fail, it's OK if the program crashes and you have to restart it. But many systems are much more constrained.
Stack usage is also much, much, much easier to characterize. In systems where stack depth is well-controlled (i.e. most embedded systems that don't have dynamic process/thread creation), very simple analysis will suffice to identify places where you blow your stack.
Not using the heap != everything is allocated on the stack. In situations where you want to avoid dynamic allocation, memory for most things that would otherwise have been dynamically allocated ends up being statically allocated at compile time.
No, but you're a lot more likely to be able to overwrite the return address via a stack-based buffer overflow, and that is generally a much more serious kind of attack.
and moves and copies won't have potential for other memory management bugs? You'll still have memory management overhead and now additional complexity.
At a high level, isn't this like implementing your own "malloc" and "free" that just pulls from your process's own memory pool instead of the OS? Or is there more to it than that?
No, it's just placing the appropriate structs and buffers on the stack (when not provided by the caller).
It does eliminate a certain couple classes of errors, and makes some others less likely.
I didn't read all the code, but I don't think it's using alloca or the like. So the stack allocation sizes are known at compile time, and bounded unless there's some recursion going on (which is unlikely).
Many real time systems and applications disallow heap usage, because they have formal verification requirements that can't be met with dynamic memory that may "run out" depending on run time state.
I love the idea of a zero allocation crypto library, but isn't the fact that this is also in C going eventually lead down a similar path as that of OpenSSL?
It's a Rust oxidization of the BoringSSL library, meaning that parts of BoringSSL are being rewritten in Rust, with the eventual goal of being pure Rust.
No, being pure Rust is not the goal. It aims to use Rust as much as possible for the parts that Rust is good at. But core crypto algorithms generally need to be written in assembler, to avoid various timing attacks that could be introduced by optimization. And for things that would require large amounts of `unsafe` in Rust, there's less reason to port that to Rust, and leaving it in C can be more clear.
The thing that are most appropriate for Rust are parsing, protocol implementation, higher level code that uses core crypto primitives, and providing a safe API to client code. But the core crypto primitives themselves will remain written in C and/or assembler, as appropriate.
> to avoid various timing attacks that could be introduced by optimization
Assembler only goes so far. Until you figure out how the processor's front end will decode the machine code and run the underlying RISC program, or how the hypervisor will schedule your program on some shared machines (e.g. in EC2) you're susceptible to a different class of side-channel attacks.
This is true but not relevant to the discussion: whatever side channel problems you'd have in pure assembly, you're practically certain to have more of them if you implement crypto in a high-level language with an aggressive optimizer.
Is there any way to write programs that bypass the processor's front end for most the processors in use? If not, it what sense is it true that "you can go further than assembler"?
This is often, though not always, the case with deeply embedded processors of the kind BearSSL seems to be designed for. There's generally some way to get predictable cycle-exact performance on them because it matters for some embedded applications.
The problem with OpenSSL is less about the language it's written in and more about the age of the project, discipline of the developers, quality of the codebase, and its prevalance - which leads to its vulnerabilities having a high impact. OpenSSL's code is a heap of trash and that's why it's vulnerable, not neccessarily because it's written in C.
Sure. Though, C does nothing to help prevent you from creating that same garbage again.
So while there are excellent examples of C projects out there, there are many more that show why it's important to provide developers (even the good ones) with guard rails.
Libraries like this are almost invariably a terrible idea: none of the more recent alternatives to OpenSSL I've seen have avoided resurrecting crypto bugs OpenSSL fixed years ago.
But: Thomas Pornin!
So, this is pretty neat. I hope lots of crypto people take a very hard look at it.
Besides being a crypto professor, he managed to guess what CRIME was about, after it was announced that some bad OpenSSL advisory was imminent, but before it came out.
Therefore the proposed bug squashing strategy of "just claim that there's a bug in XYZ and let him oracle what it is".
One of the cliches about crypto is that you should not implement your own crypto. Not to suggest that the authors don't know what they are doing, but they mention 'alpha' quality themselves on the site. I wonder, how long does it take until a new library is deemed secure? What does the process look like? Trial and error? Or do they compare notes with vulnerabilities found in e.g. openssl?
It is a combination of "many eyes" and "good documentation". What is needed is some good text that describes the design choice, the rationale, and all the tricky details; and then people who read it and think about it. I'll write and publish such text within the next few months.
Generally, open source software benefits from more users. But having a huge amount of users makes it more difficult to improve and cleanup because you can't just deprecate stuff easily. (like SSL2/3).
Also, having 100% of the internet using openssl makes the impact of a vulnerability in that library huge. Some diversity is probably a good thing.
I appreciate the time and effort that you are putting into it, good luck.
AES (Advanced Encryption Standard, 1997 to 2000): co-author of the block cipher DFC
eSTREAM (ECRYPT Stream Cipher Project, 2004 to 2008): co-author of the stream cipher SOSEMANUK (admitted in the final portfolio)
SHA-3 (2007 to 2012): co-author of the cryptographic hash function Shabal (selected for second round)
PHC (Password Hashing Competition, 2013 to 2015): author of the password hashing function Makwa (finalist, was awarded a "special recognition")
Author of the sphlib library: optimized implementations of many cryptographic hash functions, both in C and Java.
Author of RFC 6979: Deterministic Usage of the Digital Signature Algorithm (DSA) and Elliptic Curve Digital Signature Algorithm (ECDSA).
This is a very important point. When it comes to implementing new software, especially in a critical field like crypto, it takes a long time for acceptance to build.
As alpha software, I'd be shocked if anyone was using it in a production capacity, but it could be useful for early investigations into issues like timing attacks - clearly, it's better to get them sorted before a "final" release. I'd hope that any project planning to adopt this or any) crypto code was first getting it analysed carefully.
On the other hand, quite a bit of existing crypto is there because it was the first implementation on a given platform, despite potentially having issues - think about heartbleed, which was missed for quite a while in very heavily used software. It's not always bad to have fresh alternatives, as long as they are approached cautiously.
Generally the cliche is about not implementing your own cryptographic algorithms. As long as they only implement existing algorithms and don't generate new ones, I don't think this applies.
I used to think that; but I'm following Dan Boneh's crypto course on Coursera at the moment, and he specifically notes that you should not even try to implement known algorithms yourself (for production; you could do it ofcourse for the learning experience).
The reason is that there are subtle attacks on the implementation, such as timing attacks, which can leak information.
One of the point of the exercise is to provide constant-time implementations that provably do not leak such information. The hash functions, two AES, one DES/3DES, RSA and elliptic curve implementations in BearSSL have been written to be constant-time. I still have to write some document that explains how such things are done.
This thinking is why Heartbleed was such a disaster. Everyone left it to someone else, and the end result was OpenSSL being the only serious choice. Yet even the 'experts' got it wrong -- way wrong.
I'm not advocating everyone and their mother implement their own crypto. But some software diversity is a good thing. These algorithms aren't quite as scary or fickle as the documentation and existing implementations make them seem. Especially if you stick to good software like DJB's stuff.
For instance, using montgomery/edwards curves instead of weierstrass curves eliminates a lot of the difficulty in writing a constant-time implementation of ECC. And the 25519 implementation comes with a fast, constant-time implementation of a prime field type.
Yet even DJB's stuff can be simplified. You can knock off a good 80% of the scary code at a cost of a mere 10% of performance. To Google or Facebook, that may be unacceptable. But to me, that's entirely worth it. Now you have a tiny library that is easy to understand, and easy to audit.
I haven't benchmarked TweetNaCl's performance yet, but I think that goes way too far into making the code completely unreadable. The 80%/10% number I gave is against the ref10 implementation.
Please note that like BearSSL, my implementations are alpha-quality. Further, I'm not suggesting anyone use these in production. If I do so myself and it blows up in my face, it'll only have harmed me, and I'll only have myself to blame.
(Also, I'm really bad when it comes to source code comments, sorry. The why really needs you to read the research papers; the how is mostly self-evident. The remaining one-letter variable names were used to match the papers, and because I couldn't think of more descriptive terms.)
If you're using modern C++ you might consider using SaferCPlusPlus[1] for improved memory safety. Safe, fast compatible direct substitutes for C++'s unsafe elements (pointers, arrays, vectors, even references are technically unsafe). Small, easy to use, no dependency risk, and can be optionally "disabled" with a compile-time directive. It's the best answer to the Rust crowd who (justifiably) point out the inevitability of memory bugs in large enough C++ code bases.
Not sure I'd agree with that. OpenSSL is very far from perfect and obviously contains many many security bugs, but it also has a very long history of fixes, knowledge, etc. and has a large number of eyes on it. It's more of a known quantity than something new.
OpenSSL does not have half the fixes people think it does. Generally what gets fixed is the ciphersuites some people happen to use, and the rest remain broken. For example, the constant time ECDSA signing only works with some curves, not all. It still supports horrific hacks for random number generation, instead of using OS provided interfaces. NSS is not much better on this front, but does a far better job of parsing TLS records in a sane way.
There is not a single malloc() call in all the library. In fact, the whole of BearSSL requires only memcpy(), memmove(), memcmp() and strlen() from the underlying C library. This makes it utterly portable even in the most special, OS-less situations. (On “big” systems, BearSSL will automatically use a couple more system calls to access the OS-provided clock and random number generator.)
On big desktop and server OS, this feature still offers an interesting characteristic: immunity to memory leaks and memory-based DoS attacks. Outsiders cannot make BearSSL allocate megabytes of RAM since BearSSL does not actually know how to allocate RAM at all.
Whoah, no. It's about building crypto code at all.
Given a choice between building something with Nacl and a bespoke stream cipher and building something with a bespoke cryptosystem and AES, I would have a hard time picking, but I'd lean towards Nacl.
this might give people false sense of security. implementing own crypto exposes you and those few souls that bought into your story. implementing bug-free "proven crypto" is near impossible task, but "proven crypto" part sells much better and so exposes many more un-expecting victims.
My understanding is it generally takes years. First it is reviewed by multiple security professionals who look for known attacks. Once it is generally thought to be ok it sees limited real world use. There it generally just takes trial and error time.
While those algorithms all have weaknesses they are not yet completely broken and are still in wide (if declining) use. TLS 1.0 is most affected by BEAST but all modern clients have mitigations that have proven to be effective against it. The biggest issue with 3DES is its use of 64-bit blocks making it vulnerable to the SWEET32 attack, however that requires a huge amount of traffic under the same key (100s of gigabytes). SHA1 has been shown to be weak but as far as I know there are no practical attacks against it yet.
These algorithms should obviously only be used in fallback to stronger ones but they are not broken to the point where they should never be used as SSL3, RC4, and MD5 have been.
Sure, I agree -- but that's not what the page claims. It says "insecure protocol versions and choices of algorithms are not supported, by design" -- the protocols and modes that I listed are known to have various insecurities, and it still supports them.
I agree that to be useful it's necessary to support old, less secure or even insecure modes, but this is at odds with the above stated goal.
You're right technically but don't you think that's a little bit pedantic?
If your goal is to truly improve the state of the art in the ecosystem, dropping anything that is even remotely insecure is appealing I get that and I do believe the people behind BearSSL would love to do that. However to truly improve anything you need two things: Popularity and improve security.
There is a conflict there because popularity requires, at least some, compatibility to what already exists. You need to balance out security and compatibility. I think there is room for discussion about where precisely that balance is. You could further tilt it towards security by helping users of the library get a sense of what they need to support. Ultimately though you can't just blindly drop everything that's somehow not perfectly secure. Doing so would not improve security at all.
It's a small sacrifice to have one library be a little bit less secure than it could be, if that helps to make everything more secure it all.
This. I recently worked on updating an embedded TLS implementation from TLS 1.0 to TLS 1.2. I was told that it didn't need to implement TLS 1.0 or TLS 1.1, but once deployed we found a lot of non-HTTPS servers still using TLS 1.0. In particular, Microsoft's Hotmail/MSN SMTP servers and multiple RADIUS servers on WPA/WPA2 Enterprise networks. It now allows for client connections to TLS 1.0 servers, but will only serve TLS 1.2 itself.
You kind of have to support SHA-1 still. Even with the browsers moving to deprecate it, many of the root certificates valid for another 10-20 years are still using it. (since the root certs ship with the browser, the security risk is lessened.)
If this is to be a general library that validates the entire certificate chain, then you'll need SHA-1.
Now if the library tries to advertise SHA1 in ServerHello by default, then that is indeed unfortunate.
Given various clues such as:
-"OS-less"
-small memory footprint
-going out of their way to include discussion of legal jurisdictions
-the author and past activities
-performance seems not a major concern
Makes me suspect that a major goal is anonymity. It's less aimed at users installing on their non-anonymous Windows/Mac/phone but rather leverage generic/commodity hardware to communicate over SSL. Throwaway burner phones.
This just looks like someone who is aware of the complex legal history associated with cryptography and has considered embedded devices in the design of the library.
Considering the recent events around IoT having good crypto libraries for that seems like it could be useful.
From what I understand, you can use them already with AWS IoT but I'm not sure if this solution is ideal or secure enough. Haven't used it personally but it's there:
I wanted a clear situation with regards to laws on cryptographic software distribution and export. With my own server, I can keep everything in Canada, which makes things simpler.
It's hard to get by without all the fancy CVE features in GitLab, but software written in Ruby on Rails is banned in Canada, so they had to use good ol' reliable gitweb.
If Thomas has enough time to finish it (bearing in mind the hundreds of hours it will take to bring it to production quality) - and if he has time to maintain it over the years - and if he is able to commit to years' worth of future maintenance - then this could become a very nice and usable implementation.
I think it's great when people write something for the love of it.
For a high quality TLS implementation that is production-ready, and has had extensive third-party verification and certification, I'd recommend mbedTLS from ARM.
Seems like it is trying to replace PolarSSL (now called mBed TLS). It even dervies the name from PolarSSL (Polar Bear). It is nice to have multiple options, however it can make the vulnerability management a nightmare. How many more SSL libraries do we need (OpenSSL, LibreSSL,S2N,GnuTLS), not to mention native SSL libraries (Secure Transport, SChannel)?
It would probably be faster to write your own TLS library than learn all of that.
OpenSSL doesn't fare much better. So far, libtls looks the most promising. But last I checked, it was still a bit too spartan and couldn't operate in non-blocking mode, which kills you if you want an event-driven server.
I, for one, am still waiting for the TLS library that's the spiritual equivalent of NaCl or libsodium; where the integration surface is narrowed to the essentials, and sensible and secure (internal) defaults predominate.
Probably a triple entendre as the author is a frequent poster and high rep user on Security StackExchange where his profile pic has been a bear for years. The community even has a couple in jokes about it because one of the other high rep users also has a picture of a bear as his profile picture and they are affectionately referred to as "big bear" and "little bear".
Thanks, it seems the separate from variable decrease operator (q --) broke my "internal parser" as I completely missed it somehow. I'll edit my post to remove the noise.
An obvious dig at wolfSSL - I'd be curious to see a comparison between the two. wolfSSL appears to be the more mature of the two (which makes sense, it's been in use for over a decade, and has a business dedicated to its development).
Not quite, Thomas's avatar in many places is that of a bear. He has named a few of his works after bears, in particular his password hashing function Makwa.
Huh, the first thing that came to my mind when I saw the name was this thing: https://matt.ucc.asn.au/dropbear/dropbear.html . Apparently forest animals are popular in the SSL implementation writing crowd.
I'll admit, they're a local company that I have a fair bit of respect for, so I'm a bit protective about people attacking them. That said, BearSSL is just too close to be coincidence.
The author could have easily continued with their trend of using alternative languages and ended up with something just as unique. UrsidSSL, BhaalooSSL, XiongSSL; all could have continued the trend without coming across as a sly attack.
It's not as if wolfSSL is a new kid on the block in the world of embedded SSL.
It has nothing to do with WolfSSL, Thomas uses a Bear as his avatar, done. I've never heard of WolfSSL, but Thomas's name and avatar was instantly familiar as someone who's contributed a wealth of invaluable Crypto knowledge to the world.
> Honestly, all my internal git commit messages are "...".
Very insightful about how security experts write security code, thanks! All that documentation which you hope to write some time - half of it should have been in your commit messages (that's what they teach on StackOverflow, no?)
I have absolutely nothing to say when it comes to crypto, but as a C dork I found it ... quirky that the encoding/decoding functions in inner.h (https://bearssl.org/gitweb/?p=BearSSL;a=blob;f=src/inner.h;h...) seem to use e.g. uint32_t to get a 32-bit type, while assuming that char is 8 bits (rather than using uint8_t). This seems strange.
I don't think it assumes it's 8 bits, but rather it assumes the values won't be outside the range of an 8 bit number, which should be fine, given that it's an octet-oriented protocol?
Using char pointers presumably is to get correct aliasing analysis?
I've been reading the code for T0 (a FORTH->C thingy), and I arrived at the conclusion that C# is a terrible language. The code itself is nice and clean if you assume it, though.
Says who? Maybe the cruft of the old ones means its impractical to fix. Maybe the leadership of the old ones means it can't be fixed due to incompetence or toxic politics. Sometimes rolling your own or forking is the smart move. There's a reason you use x.org and not xfree86 for example.
The idea that no one should ever roll their own cryptography is a cutesey warning for amateurs, but not an absolute rule. If no one ever did, we would never have any.
Also projects like openssl don't have third-party quarterly audits or other formal practices. They're "rolling their own" as much as the other guy.
Doesn't matter anymore -- the credibility of the person leading the project is thoroughly established, so I'm retracting my comment.
But generally, "says who" is answerable as "says any reputable applied cryptographer, established audit/research team, etc. who's thoroughly cut their teeth on crypto and security in general."
Why? Crypto in a memory-managed language is not just a bad idea but a horrifically bad idea because you expose yourself to every single bug in the runtime's memory management.
Rust can't target the same platforms C can. Rust is also (generally) more memory intensive than C for similar programs. This makes its use in embedded situations a non-starter (at least for now).
I keep seeing this argument, but nobody actually names those platforms. Name one platform you need to run a TLS stack on that LLVM doesn't have support for.
> Rust is also (generally) more memory intensive than C for similar programs
As a side note, can you guarantee you are not allocating memory (calling malloc) with Rust? I know how to do it in C, but Rust has another layer of indirection.
> Name one platform you need to run a TLS stack on that LLVM doesn't have support for.
Have a look yourself:
$ llc --version # CPU architectures supported by LLVM
Of particular interest to me are the AVR ATMega line of microprocessors, popular in Arduino boards and IoT devices everywhere. I'd personally love to see more IoT communication secured with TLS.
I do see a github repo for providing a LLVM backend for AVR processors, but it's not part of the LLVM project, nor is it even passing its own test suite at the point I'm writing this.
There are so many variables here that using the benchmarks game will not give you an accurate picture. You may well just be benchmarking jemalloc vs. glibc's allocator. jemalloc is tuned for speed of allocation in threaded contexts (with e.g. TLABs), not minimum memory usage. If you want, you can use the system allocator with Rust; for the benchmarks game we use jemalloc since it's faster.
> I know how to do it in C, but Rust has another layer of indirection.
How? Any function you call from some other library can call malloc under the hood.
> $ llc --version # CPU architectures supported by LLVM
I know there are architectures out there that LLVM doesn't support, but that doesn't mean anything unless people actually need those architectures. Not having PDP-11 support (one of the architectures on that list) doesn't make Rust a "non-starter".
> I do see a github repo for providing a LLVM backend for AVR processors, but it's not part of the LLVM project, nor is it even passing its own test suite at the point I'm writing this.
AVR support is coming along well.
AVR isn't exactly a first-class citizen among commonly used projects like OpenSSL. The reason why AVR support is immature is that few people need it. Not having that support isn't enough to make Rust a "non-starter", any more than it makes OpenSSL a "non-starter".
> There are so many variables here that using the benchmarks game will not give you an accurate picture.
Hence the comment "With all the usual disclaimers about the inaccuracy of benchmarks". But it still gives a good, practical starting point for discussions. You're welcome to provide a counter data set to further the discussion.
> Any function you call from some other library can call malloc under the hood.
We're discussing low level libraries which aren't really calling out to other libraries, allowing you to quite easily control your memory allocation. It's also fairly easy to grep for 'malloc' in a C codebase, or inspect a compiled library for the malloc syscall.
I took a few minutes and looked at the current state of Rust with memory allocation as well, and it certainly does seem possible to remain within the stack only, by avoiding Box, Vec, String, and the various reference counted containers. Great to see.
There do seem to be some lingering limitations with stack-based data primitives and generics, but it's nothing you couldn't work around if you wanted to.
> unless people actually need those architectures
Like ATMega? If an architecture is supported in modern versions of GCC, there's a good chance that it's needed by somebody. And that doesn't count the dozens of specialized C compilers for other non-standard architectures. Modern banks are still running Cobol on mainframes, after all, and microcontrollers are everywhere.
> AVR support is coming along well.
Mind pointing me at a reference on, say, LLVM's roadmap? I see a fairly constant level of activity on the avr-llvm github repo, but nothing that appears to state it's ready for inclusion in LLVM in any defined timeframe.
Also curious about MSP430 moving out of experimental, Microchip's PIC, the Intel 8051, Hitachi's SuperH...
> make Rust a "non-starter", any more than it makes OpenSSL a "non-starter".
Wait, you're comparing a programming language to a heavy duty crypto library? Quite the literal version of comparing apples and oranges. Or apples and an Orange Julius.
There are other plenty of other crypto libraries that are aimed at microprocessors that aren't OpenSSL; wolfSSL as an example.
Ultimately, with enough time and money, Rust is capable of competing with C for the embedded space, making a lot of embedded developers happy. But not in the forseeable future.
> Hence the comment "With all the usual disclaimers about the inaccuracy of benchmarks". But it still gives a good, practical starting point for discussions. You're welcome to provide a counter data set to further the discussion.
There is nothing you can do in C++ that you can't do in Rust as far as memory is concerned. The compiler backends are even identical! You can drop libstd if you want in Rust, which you probably would in the microcontroller use case. You can even translate C to Rust [1], which should result in virtually identical LLVM IR!
That's what's so frustrating about throwing out benchmarks game numbers: the languages are isomorphic, so you end up ultimately comparing things like jemalloc implementations.
> Like ATMega? If an architecture is supported in modern versions of GCC, there's a good chance that it's needed by somebody. And that doesn't count the dozens of specialized C compilers for other non-standard architectures. Modern banks are still running Cobol on mainframes, after all, and microcontrollers are everywhere.
Why is COBOL on mainframes relevant? We're talking about C++ here. (Anyway, if you want to bring up mainframes, IBM has a working SystemZ backend for LLVM.)
Talking about "microcontrollers" is too broad of a brush. A lot of microcontrollers are ARM (or MIPS, etc.). Rust runs just fine on those.
> Ultimately, with enough time and money, Rust is capable of competing with C for the embedded space, making a lot of embedded developers happy. But not in the forseeable future.
This is again way too strong of a statement, because we have people using Rust right now for embedded IoT use cases (including some at Mozilla!) It all depends on what you need. If LLVM supports your architecture (which is probably does) then great! If it doesn't, then let's talk about the specific architecture you need and what you need it for, rather than making blanket "Rust is a non-starter for embedded use" statements.
We're never going to get support for every architecture anyone can come up with that C has ever run on. But who cares? What matters is whether Rust runs on a platform you were seriously considering using Rust for. If it doesn't, then we can get that fixed; chances are if you want that architecture, someone else using LLVM does too.
I don't mean "isomorphic" in that every Rust feature has a corresponding C feature. I mean that you can replicate every runtime C feature in Rust, and vice versa. The advantage of Rust is that you get a lot more safety guarantees and ergonomic features out of the compiler than you would with C.
Are any of those safety guarantees important to a crypto project, though? Even heartbleed wasn't a "buffer overrun" in the memory safety sense of that word: the client was asking for and receiving the contents of memory that the server had every right, from a safety standpoint, to read from.
Oh, wow, thank you! I totally had conflated those two in my head. (Let this go down as a rare instance of a HN commenter not doubling down on being wrong...)
our e200z6 core has PowerPC ISA, but there's a difference in hardware floating point support (here it's done by SPE) and we also need VLE. First one you could somehow get working, but with the latter - no way, its support was dropped by mainline gcc even after the patches were proposed, since it complicates the other parts of PowerPC port and is needed only by few people like me with deep embedded stuff (source: https://gcc.gnu.org/ml/gcc/2013-03/msg00172.html).
Maybe falcolas means that in the same way that C++ is more memory intensive than C for similar programs, because both C++ and Rust encourage a lot of Vecs/Strings that implicitly allocate on the heap?
Are there mature versions of rust for the hundreds of microcontrollers that people are in production right now? Most of them have a decent GCC port and C works on all of them..
In general, the versions of Rust end up being most of them, rather than being pinned to one specific version.
As my sibling mentions, if LLVM can target it, then we can, but generally anything < 32 bits isn't gonna work at the moment. We're slowly getting support for some of that stuff.
Rust uses LLVM on the backend, so it gets a lot of platform support for free. Not sure how mature that stuff is yet though. Maybe someone who's used Rust for embedded can jump in and comment?
But rust has the exact same problem: if it's "safe", it's ironically not "secure" because the algorithm no longer controls its own memory operations. If you can't say, for instance, that the key is never swapped out, then it's not secure. But if you can say the key is never swapped out, you're at a level of direct memory control that is no longer "safe".
I think "memory that is never swapped to disk" would fit nicely into Rust's ownership/borrowing model. It would probably just be a variant on Box that makes more guarantees, but which exposes the same safe interfaces? Having a safe API built on top of unsafe calls is more or less how everything in Rust works (maybe how any safe language works?), when you drill down far enough.
That said, I guess you'd have to be careful to avoid creating something like a NoSwapBox<Vec>, because the NoSwap property would only apply to the (ptr, capacity, length) triple of the vec, and not to its actual heap allocation that you probably care about. Maybe the answer there is something like requiring that the contents of NoSwapBox implement Copy, or some other way of saying "this is plain old data"? I don't know enough about the plans for custom allocators to say anything intelligent :p
you should have read the rest before commenting. This is ideal for embedded targets and C is the best way to write a truly portable library for these small targets. Consider that a lot of embedded code is designed to avoid mallocs() altogether and that helps a lot.
What would be a better language for creating a small, portable crypto library that works on embedded systems? Would you really put your trust in a brand new language?
Actually, C is a horrible language in many respects, but it is also the only language that can achieve any kind of decent compatibility in embedded systems, which is why I used it.
Also, when I say that BearSSL is written in C, it is partly a lie: some of it (especially X.509 certificate decoding, and handling of handshake messages) is done in T0, a new Forth-like language that I invented for the task (compiler is provided, and produces C code), specifically to have coroutines (incidentally, it means that most of that code read things byte by byte, with relatively few accesses to buffers, thus less potential buffer overflows).
do you really need coroutines for cert decoding and handling of handshake messages? It's not like you can parallelize these things and do something else in the meantime...
Certificates and handshake messages are nested structures, and I want to process them in a streamed fashion (e.g. I don't buffer a complete certificate, I don't have the RAM for that; BearSSL can decode certificates and messages that are larger than the total RAM is uses), so without coroutines, I would still need some sort of decoder with a stack of states, so I just accepted it and pushed the idea to its logical conclusion.
Despite the horrors implied by a Forth syntax, this actually made writing the code easier. As a bonus, it turns out that T0-generated code is very compact, so I could pack more behaviour into less bytes.
Given that this is a reasonably small code base (compared to, say, an OS kernel), an idea would be to use a language that provides strong guarantees about correctness the implementation. For example, Spark Ada can be used to do that.
Alternatively, keeping the implementation in C, a similar level of assurance could be achieved by developing the library along with a specification, and requiring any extensions to pass some kind of correctness proof based on this specification.
No really he should not. A separate translation layer is free for anyone to write though. The OpenSSL design from an API perspective is basically as far from "user friendly" as possibly possible. Having a hard-to-use API means that it's hard to get things right. If things are hard to get right it leads to more bugs. You get where I'm going with this.
A clean, simple to use, or rather hard-to-use-in-a-wrong-way API is very much needed (and there are some libraries that are nice to use but not very proven).
I don't really know what you mean about a hard-to-use api as openssl compat would just wrap the bearssl api and not make any difference to consumers of bearssl.
The problem is that OpenSSL's API itself is quite complicated and probably not in an entirely necessary way. As but one concrete example, there's almost no particularly satisfying way to handle the "error queue" in a world with imperfect software. I also recently found how how insanely impractical it is in practice to perform one's own certificate validation (e.g. so I could interpret the CN field) if a TLS connection is abstracted even a tiny bit (e.g. in a database driver).
Little doubt that someone would find it useful, but it is An Undertaking, to preserve something which is not that desirable.
> No dynamic allocation whatsoever. There is not a single malloc() call in all the library. In fact, the whole of BearSSL requires only memcpy(), memmove(), memcmp() and strlen() from the underlying C library. This makes it utterly portable even in the most special, OS-less situations. (On “big” systems, BearSSL will automatically use a couple more system calls to access the OS-provided clock and random number generator.)
> On big desktop and server OS, this feature still offers an interesting characteristic: immunity to memory leaks and memory-based DoS attacks. Outsiders cannot make BearSSL allocate megabytes of RAM since BearSSL does not actually know how to allocate RAM at all.
Edit: Just discovered what makes this an even more interesting one to watch, it's the work of this Wizard: http://security.stackexchange.com/users/655/thomas-pornin