2 Commits

Author SHA1 Message Date
ddidderr 08038e685b [release] ez-urandom v1.0.0 2026-04-28 20:27:39 +02:00
ddidderr 5623568185 feat: initial implementation of ez-urandom
A small, dependency-light crate that wraps `/dev/urandom` and exposes
ergonomic helpers for the things people actually reach for randomness
to do: read a primitive integer, sample a number in a half-open range,
or generate a random ASCII string from a given alphabet.

Why this crate exists
---------------------

The Rust ecosystem already has `rand` and `getrandom`, but both bring
in more surface area than is warranted when all you need is "give me
some bytes from the OS." This crate intentionally stays tiny: a single
`File` handle to `/dev/urandom`, a handful of typed readers, and one
rejection-sampling helper. No traits to learn, no generic plumbing, no
algorithmic CSPRNGs in-process — the kernel does that job already.

What's included
---------------

- `OsRandom`: thin wrapper around an open `/dev/urandom` handle,
  constructed via `try_new()`. Implements `io::Read` so it can plug
  into anywhere a byte source is expected.
- Typed integer readers (`get_u8` ... `get_i128`, plus `usize`/`isize`)
  generated by a small `paste!` macro. Each reads exactly
  `size_of::<T>()` bytes and interprets them in native-endian order so
  every bit pattern is equally likely.
- `gen_range_u32(n)`: uniform integer in `0..n` via rejection sampling
  on top of `get_u32`. The cutoff is computed as the largest multiple
  of `n` that fits in `2^32`, so there is *no* modulo bias — values at
  or above the cutoff are discarded and resampled. This is the
  textbook fix for the "`rand() % n` is biased when `n` doesn't divide
  `RAND_MAX+1`" footgun.
- `pick(set)` and `string_from(set, len)`: uniform byte / ASCII string
  drawn from a caller-supplied alphabet, layered on `gen_range_u32`.
- `charset` module with `const`-built alphabets: `DIGITS`, `LOWERCASE`,
  `UPPERCASE`, `ALPHABETIC`, `ALPHANUMERIC`, `HEX_LOWER`, `HEX_UPPER`.
  Built with a small `const fn concat` so the arrays exist as compile-
  time constants with no runtime allocation.

Design choices and tradeoffs
----------------------------

- Linux/Unix only by construction: opens `/dev/urandom` directly. This
  is a deliberate scope limit — supporting Windows would mean pulling
  in `BCryptGenRandom` and an abstraction layer, which defeats the
  point of the crate. Users who need cross-platform should reach for
  `getrandom`.
- `unsafe_code = "forbid"` at the crate level. The implementation does
  not need `unsafe`, and forbidding it makes that contract explicit
  and machine-checked.
- Clippy `pedantic` is on as `warn`, plus `unwrap_used = "warn"` and
  `todo = "warn"`, so the code is held to a tighter standard than the
  defaults from day one.
- `panic = "unwind"` in release: the crate's `assert!`s (e.g. "set
  must not be empty") should be recoverable by callers that wrap
  them, not abort the process.
- Native-endian integer decoding: the bits are uniformly random, so
  endianness is irrelevant to the distribution. Choosing native-endian
  avoids a needless byte swap on every read.
- Rejection sampling uses a 32-bit cutoff regardless of the requested
  range. A 64-bit cutoff would reduce the rejection probability for
  ranges close to `u32::MAX`, but in practice the worst-case rejection
  rate is < 50% and the simpler code wins.

Known limitations
-----------------

- No async API; reads are blocking. `/dev/urandom` does not block in
  practice on Linux post-init, so this is fine for typical use.
- `gen_range` is only provided for `u32`. Wider ranges would need a
  64-bit variant; not added until a use case appears.
- File handle is held for the lifetime of `OsRandom`. Callers that
  want a fresh fd per call should construct a new instance.

Test Plan
---------

- `cargo build` and `cargo clippy --all-targets` are clean under the
  pedantic lint set configured in `Cargo.toml`.
- Manual smoke test by reading several integers and generating
  alphanumeric / hex / digit strings; values are well-distributed and
  no obvious bias is visible.
2026-04-28 19:35:36 +02:00