README: rand is not a crypto library (#1514)

Closes #1358 by documenting what Rand is not.

Co-authored-by: Dan <dan.middleton@intel.com>

README.md

@@ -6,20 +6,20 @@
 [![API](https://img.shields.io/badge/api-master-yellow.svg)](https://rust-random.github.io/rand/rand)
 [![API](https://docs.rs/rand/badge.svg)](https://docs.rs/rand)
 
-Rand is a Rust library supporting random generators:
+Rand is a set of crates supporting (pseudo-)random generators:
 
--   A standard RNG trait: [`rand_core::RngCore`](https://docs.rs/rand_core/latest/rand_core/trait.RngCore.html)
+-   Built over a standard RNG trait: [`rand_core::RngCore`](https://docs.rs/rand_core/latest/rand_core/trait.RngCore.html)
--   Fast implementations of the best-in-class [cryptographic](https://rust-random.github.io/book/guide-rngs.html#cryptographically-secure-pseudo-random-number-generators-csprngs) and
+-   With fast implementations of both [strong](https://rust-random.github.io/book/guide-rngs.html#cryptographically-secure-pseudo-random-number-generators-csprngs) and
-    [non-cryptographic](https://rust-random.github.io/book/guide-rngs.html#basic-pseudo-random-number-generators-prngs) generators: [`rand::rngs`](https://docs.rs/rand/latest/rand/rngs/index.html), and more RNGs: [`rand_chacha`](https://docs.rs/rand_chacha), [`rand_xoshiro`](https://docs.rs/rand_xoshiro/), [`rand_pcg`](https://docs.rs/rand_pcg/), [rngs repo](https://github.com/rust-random/rngs/)
+    [small](https://rust-random.github.io/book/guide-rngs.html#basic-pseudo-random-number-generators-prngs) generators: [`rand::rngs`](https://docs.rs/rand/latest/rand/rngs/index.html), and more RNGs: [`rand_chacha`](https://docs.rs/rand_chacha), [`rand_xoshiro`](https://docs.rs/rand_xoshiro/), [`rand_pcg`](https://docs.rs/rand_pcg/), [rngs repo](https://github.com/rust-random/rngs/)
--   [`rand::rng`](https://docs.rs/rand/latest/rand/fn.rng.html) is an asymtotically-fast, reasonably secure generator available on all `std` targets
+-   [`rand::rng`](https://docs.rs/rand/latest/rand/fn.rng.html) is an asymptotically-fast, automatically-seeded and reasonably strong generator available on all `std` targets
--   Secure seeding via the [`getrandom` crate](https://crates.io/crates/getrandom)
+-   Direct support for seeding generators from the [`getrandom` crate](https://crates.io/crates/getrandom)
 
-Supporting random value generation and random processes:
+With broad support for random value generation and random processes:
 
--   [`Standard`](https://docs.rs/rand/latest/rand/distributions/struct.Standard.html) random value generation
+-   [`Standard`](https://docs.rs/rand/latest/rand/distributions/struct.Standard.html) random value sampling,
--   Ranged [`Uniform`](https://docs.rs/rand/latest/rand/distributions/struct.Uniform.html) number generation for many types
+    [`Uniform`](https://docs.rs/rand/latest/rand/distributions/struct.Uniform.html)-ranged value sampling
--   A flexible [`distributions`](https://docs.rs/rand/*/rand/distr/index.html) module
+    and [more](https://docs.rs/rand/latest/rand/distr/index.html)
--   Samplers for a large number of random number distributions via our own
+-   Samplers for a large number of non-uniform random number distributions via our own
     [`rand_distr`](https://docs.rs/rand_distr) and via
     the [`statrs`](https://docs.rs/statrs/0.13.0/statrs/)
 -   Random processes (mostly choose and shuffle) via [`rand::seq`](https://docs.rs/rand/latest/rand/seq/index.html) traits
@@ -28,19 +28,23 @@ All with:
 
 -   [Portably reproducible output](https://rust-random.github.io/book/portability.html)
 -   `#[no_std]` compatibility (partial)
--   *Many* performance optimisations
+-   *Many* performance optimisations thanks to contributions from the wide
+    user-base
 
-It's also worth pointing out what Rand *is not*:
+Rand **is not**:
 
--   Small. Most low-level crates are small, but the higher-level `rand` and
+-   Small (LOC). Most low-level crates are small, but the higher-level `rand`
-    `rand_distr` each contain a lot of functionality.
+    and `rand_distr` each contain a lot of functionality.
 -   Simple (implementation). We have a strong focus on correctness, speed and flexibility, but
     not simplicity. If you prefer a small-and-simple library, there are
     alternatives including [fastrand](https://crates.io/crates/fastrand)
     and [oorandom](https://crates.io/crates/oorandom).
--   Slow. We take performance seriously, with considerations also for set-up
+-   A cryptography library. Rand provides functionality for generating
-    time of new distributions, commonly-used parameters, and parameters of the
+    unpredictable random data (potentially applicable depending on requirements)
-    current sampler.
+    but does not provide high-level cryptography functionality.
+
+Rand is a community project and cannot provide legally-binding guarantees of
+security.
 
 Documentation:
 

SECURITY.md

@@ -1,31 +1,46 @@
 # Security Policy
 
-## No guarantees
+## Disclaimer
 
-Support is provided on a best-effort bases only.
+Rand is a community project and cannot provide legally-binding guarantees of
-No binding guarantees can be provided.
+security.
 
 ## Security premises
 
-Rand provides the trait `rand_core::CryptoRng` aka `rand::CryptoRng` as a marker
+### Marker traits
-trait. Generators implementing `RngCore` *and* `CryptoRng`, and given the
+
-additional constraints that:
+Rand provides the marker traits `CryptoRng`, `TryCryptoRng` and
+`CryptoBlockRng`. Generators implementing one of these traits and used in a way
+which meets the following additional constraints:
 
 -   Instances of seedable RNGs (those implementing `SeedableRng`) are
     constructed with cryptographically secure seed values
--   The state (memory) of the RNG and its seed value are not be exposed
+-   The state (memory) of the RNG and its seed value are not exposed
 
 are expected to provide the following:
 
--   An attacker can gain no advantage over chance (50% for each bit) in
+-   An attacker cannot predict the output with more accuracy than what would be
-    predicting the RNG output, even with full knowledge of all prior outputs.
+    expected through pure chance since each possible output value of any method
+    under the above traits which generates output bytes (including
+    `RngCore::next_u32`, `RngCore::next_u64`, `RngCore::fill_bytes`,
+    `TryRngCore::try_next_u32`, `TryRngCore::try_next_u64`,
+    `TryRngCore::try_fill_bytes` and `BlockRngCore::generate`) should be equally
+    likely
+-   Knowledge of prior outputs from the generator does not aid an attacker in
+    predicting future outputs
 
-For some RNGs, notably `OsRng`, `ThreadRng` and those wrapped by `ReseedingRng`,
+### Specific generators
-we provide limited mitigations against side-channel attacks:
 
--   After the state (memory) of an RNG is leaked, there is an upper-bound on the
+`OsRng` is a stateless "generator" implemented via [getrandom]. As such, it has
-    number of bits of output by the RNG before prediction of output by an
+no possible state to leak and cannot be improperly seeded.
-    observer again becomes computationally-infeasible
+
+`ThreadRng` will periodically reseed itself, thus placing an upper bound on the
+number of bits of output from an instance before any advantage an attacker may
+have gained through state-compromising side-channel attacks is lost.
+
+[getrandom]: https://crates.io/crates/getrandom
+
+### Distributions
 
 Additionally, derivations from such an RNG (including the `Rng` trait,
 implementations of the `Distribution` trait, and `seq` algorithms) should not

rand_core/src/os.rs

@@ -11,10 +11,9 @@
 use crate::{TryCryptoRng, TryRngCore};
 use getrandom::getrandom;
 
-/// A random number generator that retrieves randomness from the
+/// An interface over the operating-system's random data source
-/// operating system.
 ///
-/// This is a zero-sized struct. It can be freely constructed with `OsRng`.
+/// This is a zero-sized struct. It can be freely constructed with just `OsRng`.
 ///
 /// The implementation is provided by the [getrandom] crate. Refer to
 /// [getrandom] documentation for details.
@@ -32,7 +31,8 @@ use getrandom::getrandom;
 ///
 /// After the first successful call, it is highly unlikely that failures or
 /// significant delays will occur (although performance should be expected to
-/// be much slower than a user-space PRNG).
+/// be much slower than a user-space
+/// [PRNG](https://rust-random.github.io/book/guide-gen.html#pseudo-random-number-generators)).
 ///
 /// # Usage example
 /// ```

src/distr/uniform.rs

@@ -11,17 +11,7 @@
 //!
 //! [`Uniform`] is the standard distribution to sample uniformly from a range;
 //! e.g. `Uniform::new_inclusive(1, 6).unwrap()` can sample integers from 1 to 6, like a
-//! standard die. [`Rng::random_range`] supports any type supported by [`Uniform`].
+//! standard die. [`Rng::random_range`] is implemented over [`Uniform`].
-//!
-//! This distribution is provided with support for several primitive types
-//! (all integer and floating-point types) as well as [`std::time::Duration`],
-//! and supports extension to user-defined types via a type-specific *back-end*
-//! implementation.
-//!
-//! The types [`UniformInt`], [`UniformFloat`] and [`UniformDuration`] are the
-//! back-ends supporting sampling from primitive integer and floating-point
-//! ranges as well as from [`std::time::Duration`]; these types do not normally
-//! need to be used directly (unless implementing a derived back-end).
 //!
 //! # Example usage
 //!
@@ -151,26 +141,38 @@ use serde::{Deserialize, Serialize};
 
 /// Sample values uniformly between two bounds.
 ///
+/// # Construction
+///
 /// [`Uniform::new`] and [`Uniform::new_inclusive`] construct a uniform
-/// distribution sampling from the given range; these functions may do extra
+/// distribution sampling from the given `low` and `high` limits. `Uniform` may
-/// work up front to make sampling of multiple values faster. If only one sample
+/// also be constructed via [`TryFrom`] as in `Uniform::try_from(1..=6).unwrap()`.
-/// from the range is required, [`Rng::random_range`] can be more efficient.
+///
+/// Constructors may do extra work up front to allow faster sampling of multiple
+/// values. Where only a single sample is required it is suggested to use
+/// [`Rng::random_range`] or one of the `sample_single` methods instead.
 ///
 /// When sampling from a constant range, many calculations can happen at
 /// compile-time and all methods should be fast; for floating-point ranges and
 /// the full range of integer types, this should have comparable performance to
 /// the `Standard` distribution.
 ///
-/// Steps are taken to avoid bias, which might be present in naive
+/// # Provided implementations
-/// implementations; for example `rng.gen::<u8>() % 170` samples from the range
-/// `[0, 169]` but is twice as likely to select numbers less than 85 than other
-/// values. Further, the implementations here give more weight to the high-bits
-/// generated by the RNG than the low bits, since with some RNGs the low-bits
-/// are of lower quality than the high bits.
 ///
-/// Implementations must sample in `[low, high)` range for
+/// - `char` ([`UniformChar`]): samples a range over the implementation for `u32`
-/// `Uniform::new(low, high)`, i.e., excluding `high`. In particular, care must
+/// - `f32`, `f64` ([`UniformFloat`]): samples approximately uniformly within a
-/// be taken to ensure that rounding never results values `< low` or `>= high`.
+///   range; bias may be present in the least-significant bit of the significand
+///   and the limits of the input range may be sampled even when an open
+///   (exclusive) range is used
+/// - Integer types ([`UniformInt`]) may show a small bias relative to the
+///   expected uniform distribution of output. In the worst case, bias affects
+///   1 in `2^n` samples where n is 56 (`i8` and `u8`), 48 (`i16` and `u16`), 96
+///   (`i32` and `u32`), 64 (`i64` and `u64`), 128 (`i128` and `u128`).
+///   The `unbiased` feature flag fixes this bias.
+/// - `usize` ([`UniformUsize`]) is handled specially, using the `u32`
+///   implementation where possible to enable portable results across 32-bit and
+///   64-bit CPU architectures.
+/// - `Duration` ([`UniformDuration`]): samples a range over the implementation
+///   for `u32` or `u64`
 ///
 /// # Example
 ///

src/distr/uniform_float.rs

@@ -29,14 +29,17 @@ use serde::{Deserialize, Serialize};
 ///
 /// # Implementation notes
 ///
-/// Instead of generating a float in the `[0, 1)` range using [`Standard`], the
+/// `UniformFloat` implementations convert RNG output to a float in the range
-/// `UniformFloat` implementation converts the output of an PRNG itself. This
+/// `[1, 2)` via transmutation, map this to `[0, 1)`, then scale and translate
-/// way one or two steps can be optimized out.
+/// to the desired range. Values produced this way have what equals 23 bits of
+/// random digits for an `f32` and 52 for an `f64`.
 ///
-/// The floats are first converted to a value in the `[1, 2)` interval using a
+/// # Bias and range errors
-/// transmute-based method, and then mapped to the expected range with a
+///
-/// multiply and addition. Values produced this way have what equals 23 bits of
+/// Bias may be expected within the least-significant bit of the significand.
-/// random digits for an `f32`, and 52 for an `f64`.
+/// It is not guaranteed that exclusive limits of a range are respected; i.e.
+/// when sampling the range `[a, b)` it is not guaranteed that `b` is never
+/// sampled.
 ///
 /// [`new`]: UniformSampler::new
 /// [`new_inclusive`]: UniformSampler::new_inclusive

src/distr/uniform_int.rs

@@ -58,6 +58,13 @@ use serde::{Deserialize, Serialize};
 /// multiply by `range`, the result is in the high word. Then comparing the low
 /// word against `zone` makes sure our distribution is uniform.
 ///
+/// # Bias
+///
+/// Unless the `unbiased` feature flag is used, outputs may have a small bias.
+/// In the worst case, bias affects 1 in `2^n` samples where n is
+/// 56 (`i8` and `u8`), 48 (`i16` and `u16`), 96 (`i32` and `u32`), 64 (`i64`
+/// and `u64`), 128 (`i128` and `u128`).
+///
 /// [`Uniform`]: super::Uniform
 #[derive(Clone, Copy, Debug, PartialEq, Eq)]
 #[cfg_attr(feature = "serde", derive(Serialize, Deserialize))]

src/rngs/thread.rs

@@ -41,16 +41,36 @@ const THREAD_RNG_RESEED_THRESHOLD: u64 = 1024 * 64;
 /// A reference to the thread-local generator
 ///
 /// This type is a reference to a lazily-initialized thread-local generator.
-/// An instance can be obtained via [`rand::rng()`][crate::rng())] or via
+/// An instance can be obtained via [`rand::rng()`][crate::rng()] or via
-/// `ThreadRng::default()`.
+/// [`ThreadRng::default()`].
 /// The handle cannot be passed between threads (is not `Send` or `Sync`).
 ///
-/// `ThreadRng` uses the same CSPRNG as [`StdRng`], ChaCha12. As with
+/// # Security
-/// [`StdRng`], the algorithm may be changed, subject to reasonable expectations
-/// of security and performance.
 ///
-/// `ThreadRng` is automatically seeded from [`OsRng`] with periodic reseeding
+/// Security must be considered relative to a threat model and validation
-/// (every 64 kiB — see [`ReseedingRng`] documentation for details).
+/// requirements. The Rand project can provide no guarantee of fitness for
+/// purpose. The design criteria for `ThreadRng` are as follows:
+///
+/// - Automatic seeding via [`OsRng`] and periodically thereafter (see
+///   ([`ReseedingRng`] documentation). Limitation: there is no automatic
+///   reseeding on process fork (see [below](#fork)).
+/// - A rigorusly analyzed, unpredictable (cryptographic) pseudo-random generator
+///   (see [the book on security](https://rust-random.github.io/book/guide-rngs.html#security)).
+///   The currently selected algorithm is ChaCha (12-rounds).
+///   See also [`StdRng`] documentation.
+/// - Not to leak internal state through [`Debug`] or serialization
+///   implementations.
+/// - No further protections exist to in-memory state. In particular, the
+///   implementation is not required to zero memory on exit (of the process or
+///   thread). (This may change in the future.)
+/// - Be fast enough for general-purpose usage. Note in particular that
+///   `ThreadRng` is designed to be a "fast, reasonably secure generator"
+///   (where "reasonably secure" implies the above criteria).
+///
+/// We leave it to the user to determine whether this generator meets their
+/// security requirements. For an alternative, see [`OsRng`].
+///
+/// # Fork
 ///
 /// `ThreadRng` is not automatically reseeded on fork. It is recommended to
 /// explicitly call [`ThreadRng::reseed`] immediately after a fork, for example:
@@ -68,13 +88,6 @@ const THREAD_RNG_RESEED_THRESHOLD: u64 = 1024 * 64;
 /// from an interrupt (e.g. a fork handler) unless it can be guaranteed that no
 /// other method on the same `ThreadRng` is currently executing.
 ///
-/// Security must be considered relative to a threat model and validation
-/// requirements. `ThreadRng` attempts to meet basic security considerations
-/// for producing unpredictable random numbers: use a CSPRNG, use a
-/// recommended platform-specific seed ([`OsRng`]), and avoid
-/// leaking internal secrets e.g. via [`Debug`] implementation or serialization.
-/// Memory is not zeroized on drop.
-///
 /// [`ReseedingRng`]: crate::rngs::ReseedingRng
 /// [`StdRng`]: crate::rngs::StdRng
 #[derive(Clone)]
@@ -115,9 +128,9 @@ thread_local!(
     }
 );
 
-/// Access a local, pre-initialized generator
+/// Access a fast, pre-initialized generator
 ///
-/// This is a reasonably fast unpredictable thread-local instance of [`ThreadRng`].
+/// This is a handle to the local [`ThreadRng`].
 ///
 /// See also [`crate::rngs`] for alternatives.
 ///
@@ -139,6 +152,10 @@ thread_local!(
 /// println!("A simulated die roll: {}", rng.random_range(1..=6));
 /// # }
 /// ```
+///
+/// # Security
+///
+/// Refer to [`ThreadRng#Security`].
 pub fn rng() -> ThreadRng {
     let rng = THREAD_RNG_KEY.with(|t| t.clone());
     ThreadRng { rng }