[weland]

It doesn't have custom allocator support as in "you can't have one function allocate memory and pass it for another function to use it", or as in "you can't replace the runtime's own malloc"? OpenSSL were doing the former, not the latter.

(Edit: I'm really really curious, not necessarily trying to prove a point. I deal with low-level code in safety-critical (think medical) stuff every day, and only lack of time is what makes me procrastinate that week when I'm finally going to learn Rust)

[markus2012]

Well, anything is possible because ... human ingenuity.

However, Rust currently statically links in jemalloc - even when building a dynamic shared library. There is no easy way around it.

(because someone might ask: rustc -C prefer-dynamic dynamically links everything except jemalloc)

Having said that, I hope jemalloc gets linked externally soon so my code doesn't have to pay the memory penalty in each of my memory-constrained threads.

[steveklabnik]

You can't say "this vector uses this allocator and this vector uses another one." If you throw away the standard library, you can implement malloc yourself, but then, you're building all of your own stuff on top of it, so you'd be in control of whatever in that case.

(We eventually plan on supporting this case, just haven't gotten there yet.)

[weland]

Yes, my point was that OpenSSL did not throw away the standard library! openssl_{malloc|free} were thin wrappers over the native malloc() and free(), except that they tried to be clever and not natively free() memory areas so that they can be reused by openssl_malloc() without calling malloc() again. I.e. sometimes, openssl_free(buf) would not call free(buf), but leave buf untouched and put it in a queue -- and the openssl_malloc() would take it out of there and give it to callers.

So basically openssl_malloc() wasn't always malloc()-ing -- it would sometimes return a previously malloc()-ed aread that was never (actually) freed.

This rendered OpenBSD's hardened malloc() useless: you can configure OBSD to always initialize malloc-ed areas. If OpenSSL had used malloc() instead of openssl_malloc, even with the wrong (large and unchecked) size, the buffer would not have contained previous values, as it would have already been initialized to 0x0D. Instead, they'd return previous malloc()-ed -- and never actually free()-d -- buffers, that contained the old data. Since malloc() was not called for them, there was never a chance to re-initialize their contents.

This can trivially (if just as dumbly!) be implemented in Go. It's equally pointless (the reasons they did that were 100% historical), but possible. And -- assuming they'd have done away with the whole openssl_dothis() and openssl_dothat() crap -- heartbleed would have been trivially prevented in C by just sanely implementing malloc.

> (We eventually plan on supporting this case, just haven't gotten there yet.)

I'm really (and not maliciously!) curious about how this will be done :). You guys are doing some amazing work with Rust! Good luck!

[steveklabnik]

https://github.com/rust-lang/rfcs/issues/538 is where we're tracking the discussion, basically :) And thanks!