Random number generator updates for Linux 5.17
lore.kernel.org
lore.kernel.org
The thing is, with Linux development, these things tend to work their way in _gradually_, rather than totally radical rewrites. (If you recall, I tried the radical rewrite thing with my Zinc replacement [1] for the crypto API, which wasn't too well received.) So I'm definitely not looking to cause a ruckus, but I do really want to get this into good shape. Provably secure constructions and slow-but-steady work will hopefully be key in doing that.
So, that may not the concrete answer perhaps you were hoping for, but that's a snapshot of my general approach.
The use of RDRAND was especially overzealous. The problem is always finding the line between retaining entropy and performance - in this case (without having done the statistics) I believe you have done the right thing.
If I wanted to contribute an entropy mixer not based on CRC32 (as the current one is) that is much faster, includes test vectors and entropy analysis - where would be a good place to start?
A random device appeared in FreeBSD 2.2. The implementation was changed
to the Yarrow algorithm in FreeBSD 5.0. In FreeBSD 11.0, the Fortuna al-
gorithm was introduced as the default. In FreeBSD 12.0, Yarrow was re-
moved entirely.
* https://www.schneier.com/academic/fortuna/
Also a video presentation "A Deep Dive into FreeBSD's Kernel RNG" from vBSD 2017:
* https://papers.freebsd.org/2017/vbsdcon/gurney-a_deep_dive_i...
I must say (and I say this very much as a rabid Linux fanboy) that it bothered me a bit how Linux's random number generator for a long time did not appear to be developed to a particular published and reviewed specification, but rather it seemed to be a somewhat ad hoc affair. This might be completely mischaracterizing it unfairly, but that's just how it comes across.
Does Linux's latest RNG design (which people say is finally good) conform to any mature and well analyzed and characterized published standard? I'm not asking if it uses standardized ciphers in particular stages, but whether the entire RNG follows a standard.
(Why not use Fortuna as-written? The main reason is that it is not designed to scale. It's a reasonable design for the single core machines of the era when it was written, but it does not worry about contention at all. You could also swap out AES for Chacha as PRF, because: Chacha is faster to seed, faster to generate on machines without AES-NI, and doesn't have some of the theoretical problems -- discussed in the Fortuna chapter -- that AES has with its relatively small 128-bit blocks.)
random(4): Fortuna: allow increased concurrency
Add experimental feature to increase concurrency in Fortuna. As this
diverges slightly from canonical Fortuna, and due to the security
sensitivity of random(4), it is off by default. To enable it, set the
tunable kern.random.fortuna.concurrent_read="1". The rest of this commit
message describes the behavior when enabled.
[…]
random(4): Flip default Fortuna generator over to Chacha20
The implementation was landed in r344913 and has had some bake time (at
least on my personal systems). There is some discussion of the motivation
for defaulting to this cipher as a PRF in the commit log for r344913.
As documented in that commit, administrators can retain the prior (AES-ICM)
mode of operation by setting the 'kern.random.use_chacha20_cipher' tunable
to 0 in loader.conf(5).
Approved by: csprng(delphij, markm)
Differential Revision: https://reviews.freebsd.org/D22878
There are standard RNG designs that NIST publishes, but they're low level, on the order of constructions, not whole system designs. A lot of the important details are OS-specific. It shouldn't make you more comfortable to hear that a system uses something like HMAC_DRBG; that detail doesn't tell you a whole lot about how the system as a whole works.
† See, for instance: https://eprint.iacr.org/2013/338.pdf --- skip to Section 5 for a pretty detailed description of the LRNG.
random: try to actively add entropy rather than passively wait for it
random: only read from /dev/random after its pool has received 128 bits
random: Return nbytes filled from hw RNG
random: mix rdrand with entropy sent in from userspace
random: use a different mixing algorithm for add_device_randomness()
random: add backtracking protection to the CRNG
random: replace non-blocking pool with a Chacha20-based CRNG
random: use an improved fast_mix() function
random: cap the rate which the /dev/urandom pool gets reseeded
random: account for entropy loss due to overwrites
random: allow fractional bits to be tracked
random: replace non-blocking pool with a Chacha20-based CRNG
The CRNG is faster, and we don't pretend to track entropy usage in the
CRNG any more.
Well that's basically what a standard is, in my mind. I place little weight on additional initials of organizations which publish standards, I just think they should be there so the design can be analyzed and examined as a whole, and by people who are not experts in C programming or want to wade through the Linux source.
The section of the paper you linked would be a fine standard for Linux's RNG except that it's a continually moving target.
And a complete design specification can certainly address the practical details of integration into a kernel, the random number subsystem in Linux has just a small interface with the rest of the kernel that can easily be captured abstractly.
1. It's not an unalloyed good thing to use "Fortuna", which is not so much a peer-reviewed standard as "a thing in a book with Bruce Schneier's name on it". I might be a little more nervous about a design that advertises being based on "Fortuna" than a random design, because the hard parts of doing a kernel CSPRNG are I think mostly not what they wrote about in that book.
2. Whether or not you have a reference design to base off of, you're still going to end up with a stream of fiddly commits determining i.e. when the generator is seeded well enough to release random bits to callers, or the order in which raw events are fed into the generator. Those annoying fiddly bits are the challenge of maintaining a secure RNG, and there's no standard anywhere that will release you from having to pay attention to that stuff (thankfully, Jason Donenfeld is doing that thankless work for us now).
The rest of it, sure. The LRNG sucks, and is ill-specified (if relatively well studied), and has a janky history. It's in good hands now! It should not be rewritten to comply with some "standard"! But sure, the rest of your arguments, well taken, &c &c.
* FreeBSD generation (getrandom(), /dev/urandom, arc4random, etc) does not use RDRAND in the hot path.
* The Fortuna state hashing algorithm uses a SHA256-based construction; not SHA1 or Blake2s[1].
[1]: https://github.com/freebsd/freebsd-src/blob/main/sys/dev/ran... , https://github.com/freebsd/freebsd-src/blob/main/sys/dev/ran...
Totally negligible. It speeds up benchmarks of getrandom() and /dev/urandom, which is a good thing, but if you were spending any significant CPU there, it was due to an architectural problem that this doesn't solve.
> It seems that, in addition to an overdue replacement SHA-1, the new code is a significant increase in performance.
No, the big improvement is removing RDRAND (arch_get_random_long) from the hot path, which only existed on relatively fast x86 hardware. (The only other arch with a non-empty arch_get_random_long implementation was s390, a mainframe.) SHA1 to Blake is only used periodically (~0.003 Hz) and probably not noticeable.
I don't think it is hard to imagine an increase in throughput for disk encryption due to less time spent in the RNG.