diff options
author | Brian Olsen <brian@maven-group.org> | 2024-07-22T13·44+0200 |
---|---|---|
committer | clbot <clbot@tvl.fyi> | 2024-08-16T16·26+0000 |
commit | 3cd57ce4e318a6126322d3298305c822acee08f0 (patch) | |
tree | 032e4209638428b731c0c476205274f546fc9cf5 /tvix/glue | |
parent | 9d2017624136e20b97dbe7cae1d9b11152f0c8d9 (diff) |
feat(tvix/glue): Add refscanner pattern and AsyncRead r/8502
This splits the existing ReferenceScanner into a ReferenceScanner and ReferencePattern as well as adds an AsyncRead implementation that can do a scan while you read from it. The reason to split the scanner in two is that generating the pattern is expensive and when ingesting build results with multiple outputs you want to do several independant scans that look for the same pattern. The reader is for scanning files without having to load the entire file into memory. Change-Id: I993f5a32308c12d9035840f8e04fe82e8dc1d962 Reviewed-on: https://cl.tvl.fyi/c/depot/+/12052 Autosubmit: Brian Olsen <me@griff.name> Tested-by: BuildkiteCI Reviewed-by: flokli <flokli@flokli.de>
Diffstat (limited to 'tvix/glue')
-rw-r--r-- | tvix/glue/Cargo.toml | 1 | ||||
-rw-r--r-- | tvix/glue/src/refscan.rs | 265 |
2 files changed, 244 insertions, 22 deletions
diff --git a/tvix/glue/Cargo.toml b/tvix/glue/Cargo.toml index 6e9f313aa0a9..7784e08f523c 100644 --- a/tvix/glue/Cargo.toml +++ b/tvix/glue/Cargo.toml @@ -45,6 +45,7 @@ nix = { version = "0.27.1", features = [ "fs" ] } pretty_assertions = "1.4.0" rstest = "0.19.0" tempfile = "3.8.1" +tokio-test = "0.4.3" [features] default = ["nix_tests"] diff --git a/tvix/glue/src/refscan.rs b/tvix/glue/src/refscan.rs index 0e0bb6c77828..4393c98614e5 100644 --- a/tvix/glue/src/refscan.rs +++ b/tvix/glue/src/refscan.rs @@ -6,62 +6,247 @@ //! //! The scanner itself is using the Wu-Manber string-matching algorithm, using //! our fork of the `wu-mamber` crate. - +use pin_project::pin_project; use std::collections::BTreeSet; +use std::pin::Pin; +use std::sync::Arc; +use std::task::{ready, Poll}; +use tokio::io::{AsyncBufRead, AsyncRead, ReadBuf}; use wu_manber::TwoByteWM; -pub const STORE_PATH_LEN: usize = "/nix/store/00000000000000000000000000000000".len(); - -/// Represents a "primed" reference scanner with an automaton that knows the set -/// of store paths to scan for. -pub struct ReferenceScanner<P: Ord + AsRef<[u8]>> { +/// A searcher that incapsulates the candidates and the Wu-Manber searcher. +/// This is separate from the scanner because we need to look for the same +/// pattern in multiple outputs and don't want to pay the price of constructing +/// the searcher for each build output. +pub struct ReferencePatternInner<P> { candidates: Vec<P>, + longest_candidate: usize, + // FUTUREWORK: Support overlapping patterns to be compatible with cpp Nix searcher: Option<TwoByteWM>, - matches: Vec<usize>, } -impl<P: Clone + Ord + AsRef<[u8]>> ReferenceScanner<P> { - /// Construct a new `ReferenceScanner` that knows how to scan for the given - /// candidate store paths. +#[derive(Clone)] +pub struct ReferencePattern<P> { + inner: Arc<ReferencePatternInner<P>>, +} + +impl<P> ReferencePattern<P> { + pub fn candidates(&self) -> &[P] { + &self.inner.candidates + } + + pub fn longest_candidate(&self) -> usize { + self.inner.longest_candidate + } +} + +impl<P: AsRef<[u8]>> ReferencePattern<P> { + /// Construct a new `ReferencePattern` that knows how to scan for the given + /// candidates. pub fn new(candidates: Vec<P>) -> Self { let searcher = if candidates.is_empty() { None } else { Some(TwoByteWM::new(&candidates)) }; + let longest_candidate = candidates.iter().fold(0, |v, c| v.max(c.as_ref().len())); - ReferenceScanner { - searcher, - candidates, - matches: Default::default(), + ReferencePattern { + inner: Arc::new(ReferencePatternInner { + searcher, + candidates, + longest_candidate, + }), } } +} + +impl<P> From<Vec<P>> for ReferencePattern<P> +where + P: AsRef<[u8]>, +{ + fn from(candidates: Vec<P>) -> Self { + Self::new(candidates) + } +} + +/// Represents a "primed" reference scanner with an automaton that knows the set +/// of bytes patterns to scan for. +pub struct ReferenceScanner<P> { + pattern: ReferencePattern<P>, + matches: Vec<bool>, +} - /// Scan the given str for all non-overlapping matches and collect them +impl<P: AsRef<[u8]>> ReferenceScanner<P> { + /// Construct a new `ReferenceScanner` that knows how to scan for the given + /// candidate bytes patterns. + pub fn new<IP: Into<ReferencePattern<P>>>(pattern: IP) -> Self { + let pattern = pattern.into(); + let matches = vec![false; pattern.candidates().len()]; + ReferenceScanner { pattern, matches } + } + + /// Scan the given buffer for all non-overlapping matches and collect them /// in the scanner. pub fn scan<S: AsRef<[u8]>>(&mut self, haystack: S) { - if haystack.as_ref().len() < STORE_PATH_LEN { + if haystack.as_ref().len() < self.pattern.longest_candidate() { return; } - if let Some(searcher) = &self.searcher { + if let Some(searcher) = &self.pattern.inner.searcher { for m in searcher.find(haystack) { - self.matches.push(m.pat_idx); + self.matches[m.pat_idx] = true; } } } + pub fn pattern(&self) -> &ReferencePattern<P> { + &self.pattern + } + + pub fn matches(&self) -> &[bool] { + &self.matches + } + + pub fn candidate_matches(&self) -> impl Iterator<Item = &P> { + let candidates = self.pattern.candidates(); + self.matches.iter().enumerate().filter_map(|(idx, found)| { + if *found { + Some(&candidates[idx]) + } else { + None + } + }) + } +} + +impl<P: Clone + Ord + AsRef<[u8]>> ReferenceScanner<P> { /// Finalise the reference scanner and return the resulting matches. pub fn finalise(self) -> BTreeSet<P> { - self.matches - .into_iter() - .map(|idx| self.candidates[idx].clone()) - .collect() + self.candidate_matches().cloned().collect() + } +} + +const DEFAULT_BUF_SIZE: usize = 8 * 1024; + +#[pin_project] +pub struct ReferenceReader<P, R> { + scanner: ReferenceScanner<P>, + buffer: Vec<u8>, + consumed: usize, + #[pin] + reader: R, +} + +impl<P, R> ReferenceReader<P, R> +where + P: AsRef<[u8]>, +{ + pub fn new(pattern: ReferencePattern<P>, reader: R) -> ReferenceReader<P, R> { + Self::with_capacity(DEFAULT_BUF_SIZE, pattern, reader) + } + + pub fn with_capacity( + capacity: usize, + pattern: ReferencePattern<P>, + reader: R, + ) -> ReferenceReader<P, R> { + // If capacity is not at least as long as longest_candidate we can't do a scan + let capacity = capacity.max(pattern.longest_candidate()); + ReferenceReader { + scanner: ReferenceScanner::new(pattern), + buffer: Vec::with_capacity(capacity), + consumed: 0, + reader, + } + } + + pub fn scanner(&self) -> &ReferenceScanner<P> { + &self.scanner + } +} + +impl<P, R> ReferenceReader<P, R> +where + P: Clone + Ord + AsRef<[u8]>, +{ + pub fn finalise(self) -> BTreeSet<P> { + self.scanner.finalise() + } +} + +impl<P, R> AsyncRead for ReferenceReader<P, R> +where + R: AsyncRead, + P: AsRef<[u8]>, +{ + fn poll_read( + mut self: Pin<&mut Self>, + cx: &mut std::task::Context<'_>, + buf: &mut tokio::io::ReadBuf<'_>, + ) -> Poll<std::io::Result<()>> { + let internal_buf = ready!(self.as_mut().poll_fill_buf(cx))?; + let amt = buf.remaining().min(internal_buf.len()); + buf.put_slice(&internal_buf[..amt]); + self.consume(amt); + Poll::Ready(Ok(())) + } +} + +impl<P, R> AsyncBufRead for ReferenceReader<P, R> +where + R: AsyncRead, + P: AsRef<[u8]>, +{ + fn poll_fill_buf( + self: Pin<&mut Self>, + cx: &mut std::task::Context<'_>, + ) -> Poll<std::io::Result<&[u8]>> { + let overlap = self.scanner.pattern.longest_candidate() - 1; + let mut this = self.project(); + // Still data in buffer + if *this.consumed < this.buffer.len() { + return Poll::Ready(Ok(&this.buffer[*this.consumed..])); + } + // We need to copy last `overlap` bytes to front to deal with references that overlap reads + if *this.consumed > overlap { + let start = this.buffer.len() - overlap; + this.buffer.copy_within(start.., 0); + this.buffer.truncate(overlap); + *this.consumed = overlap; + } + // Read at least until self.buffer.len() > overlap so we can do one scan + loop { + let filled = { + let mut buf = ReadBuf::uninit(this.buffer.spare_capacity_mut()); + ready!(this.reader.as_mut().poll_read(cx, &mut buf))?; + buf.filled().len() + }; + // SAFETY: We just read `filled` amount of data above + unsafe { + this.buffer.set_len(filled + this.buffer.len()); + } + if filled == 0 || this.buffer.len() > overlap { + break; + } + } + this.scanner.scan(&this.buffer); + Poll::Ready(Ok(&this.buffer[*this.consumed..])) + } + + fn consume(self: Pin<&mut Self>, amt: usize) { + debug_assert!(self.consumed + amt <= self.buffer.len()); + let this = self.project(); + *this.consumed += amt; } } #[cfg(test)] mod tests { + use rstest::rstest; + use tokio::io::AsyncReadExt as _; + use tokio_test::io::Builder; + use super::*; // The actual derivation of `nixpkgs.hello`. @@ -112,4 +297,40 @@ mod tests { assert!(result.contains(c)); } } + + #[rstest] + #[case::normal(8096, 8096)] + #[case::small_capacity(8096, 1)] + #[case::small_read(1, 8096)] + #[case::all_small(1, 1)] + #[tokio::test] + async fn test_reference_reader(#[case] chunk_size: usize, #[case] capacity: usize) { + let candidates = vec![ + // these exist in the drv: + "33l4p0pn0mybmqzaxfkpppyh7vx1c74p", + "pf80kikyxr63wrw56k00i1kw6ba76qik", + "cp65c8nk29qq5cl1wyy5qyw103cwmax7", + // this doesn't: + "fn7zvafq26f0c8b17brs7s95s10ibfzs", + ]; + let pattern = ReferencePattern::new(candidates.clone()); + let mut mock = Builder::new(); + for c in HELLO_DRV.as_bytes().chunks(chunk_size) { + mock.read(c); + } + let mock = mock.build(); + let mut reader = ReferenceReader::with_capacity(capacity, pattern, mock); + let mut s = String::new(); + reader.read_to_string(&mut s).await.unwrap(); + assert_eq!(s, HELLO_DRV); + + let result = reader.finalise(); + assert_eq!(result.len(), 3); + + for c in candidates[..3].iter() { + assert!(result.contains(c)); + } + } + + // FUTUREWORK: Test with large file } |