0.18 round 2: refactor Disc::patch onto Pipeline + PatchSink

Patch was strictly serial (per-sector recovery: read → seek+write
→ mapfile.record → next). Lifting the write+record onto a consumer
thread lets the drive issue the next per-sector retry while the
previous block's recovered bytes are being committed — small but
real win on damaged discs with many bad sectors, and uniform with
sweep's threading model.

- New PatchSink: Sink<PatchItem> impl in src/disc/patch.rs. Owns
  WritebackFile + Mapfile. apply() seeks+writes recovered bytes
  and records mapfile state per item; close() runs sync_all and
  mapfile.flush.
- Channel depth: WRITE_THROUGH_DEPTH (1). Patch wants minimum
  buffering — back-pressure should kick in immediately so the
  drive's per-sector retry budget isn't ahead of the writer.
- Disc::patch: keeps every existing recovery decision on the
  producer (reverse walk, damage-window skip, NOT_READY pauses,
  bridge-degradation handling, wedge exit, range watchdog).
  WritebackFile ownership moves to the sink.

Behaviour-preserving: per-sector single-shot read budget unchanged
(BU40N+Initio bridge wedge concern still respected); recovery
algorithm bit-identical.

See freemkv-private/memory/0_18_redesign.md.

Single contributor: MattJackson.
This commit is contained in:
2026-05-09 10:28:14 -07:00
parent 0cd7314831
commit 3f486b6aa8
5 changed files with 527 additions and 139 deletions
+241
View File
@@ -0,0 +1,241 @@
//! Producer / consumer split for `Disc::patch`.
//!
//! Background: pre-0.18 patch ran strictly serial — single-sector
//! recovery read → seek + write recovered bytes → mapfile.record →
//! next iteration. The drive sat idle while the previous block's
//! recovered bytes were committed. On a damaged disc with many bad
//! sectors that adds up: per-sector write + mapfile.record costs a
//! handful of milliseconds each, which the drive could be using to
//! issue the next per-sector retry.
//!
//! This module decouples them. A consumer thread owns the
//! [`crate::io::WritebackFile`] (the ISO file) and the
//! [`super::mapfile::Mapfile`]. The producer thread (`Disc::patch`)
//! keeps the [`crate::sector::SectorReader`], the wedge / damage-window
//! state, the per-range watchdog, decrypt — so what enters the channel
//! is already-clean cleartext bytes (or an "Unreadable" terminal mark).
//!
//! Producer and consumer run concurrently; the channel uses
//! [`crate::io::pipeline::WRITE_THROUGH_DEPTH`] (=1) so back-pressure
//! kicks in immediately. We want the drive's per-sector retry budget
//! to stay in lockstep with the writer — sweep's `DEFAULT_PIPELINE_DEPTH`
//! (4) would let several sectors of recovered bytes queue up between
//! the producer's retry decisions and the writer, and patch's recovery
//! loop reads stats (`bytes_good`, range progress) inline to drive its
//! skip / wedge decisions. WRITE_THROUGH_DEPTH gives "read N+1 while
//! writing N", no further pipelining — exactly the model the producer
//! logic was written against.
//!
//! Correctness invariants preserved:
//! - Mapfile is single-writer (consumer-only). No locking on it.
//! - All recovery state (damage window, consecutive_failures, skip
//! escalation, range watchdog) stays on the producer thread.
//! - `set_speed` calls happen on the producer thread (same thread that
//! owns the `SectorReader`). No new SCSI concurrency.
//! - Per-iteration ordering of file-write → mapfile-record is kept
//! intact in the consumer (write before record), so the on-disk
//! invariant "mapfile only marks Finished what the file has received"
//! survives a crash mid-pass.
//! - The BU40N+Initio bridge wedge concern is unchanged: only one
//! SCSI command in flight at a time, error-path timing identical,
//! no new retry logic. The threading primitive only overlaps the
//! *write* with the *next read*; the per-sector single-shot read
//! budget that the bridge wedge concern was originally about is
//! untouched.
//!
//! Per-range watchdog (`MAX_RANGE_SECS` / `RANGE_BUDGET_CAP_SECS`)
//! checks `bytes_good` for forward progress. With work in flight on
//! the consumer, the producer would otherwise see stale values; the
//! sink publishes a [`SharedPatchState`] snapshot after every record
//! so the producer's stall guards observe consumer side-effects with
//! at most one item of lag (which is fine — the watchdog uses minute-
//! scale budgets, not single-record latency).
use std::io::{Seek, SeekFrom, Write};
use std::sync::{Arc, Mutex};
use crate::error::{Error, Result};
use crate::io::pipeline::{Flow, Sink};
use super::mapfile::{MapStats, Mapfile, SectorStatus};
/// Item the producer hands to the patch consumer. One per per-sector
/// recovery decision.
pub(super) enum PatchItem {
/// Sector / small batch successfully recovered (and decrypted on the
/// producer side if `opts.decrypt` was set). Consumer seeks to
/// `pos`, writes `buf`, records the range as `Finished`.
Recovered { pos: u64, buf: Vec<u8> },
/// Producer exhausted retries on `[pos, pos+len)`. Consumer records
/// the range as `Unreadable`. No file write — the existing zero-fill
/// from sweep is preserved in place.
Unreadable { pos: u64, len: u64 },
/// Producer hit the per-range skip limit and is leaving the
/// remaining bytes as `NonTrimmed` for a future pass. CRITICAL:
/// this is not the same as `Unreadable` — sectors we never tried
/// stay hopeful. (See the comment at the skip-limit branch in
/// `Disc::patch`: ~36% of patch-marked Unreadable sectors are
/// actually readable on a later pass.) No file write.
NonTrimmed { pos: u64, len: u64 },
}
/// Mapfile snapshot the sink republishes after every record so the
/// producer can drive its stall / progress logic without holding the
/// mapfile lock for long. `bad_ranges` mirrors what
/// `Mapfile::ranges_with(&[NonTrimmed, Unreadable, NonScraped, NonTried])`
/// would return — same set the pre-split patch loop computed inline
/// for the progress callback.
pub(super) struct SharedPatchState {
pub stats: MapStats,
pub bad_ranges: Vec<(u64, u64)>,
}
impl SharedPatchState {
fn from_map(map: &Mapfile) -> Self {
Self {
stats: map.stats(),
bad_ranges: map.ranges_with(&[
SectorStatus::NonTrimmed,
SectorStatus::Unreadable,
SectorStatus::NonScraped,
SectorStatus::NonTried,
]),
}
}
}
/// Final summary returned by [`Sink::close`] when the consumer drains
/// cleanly. Mirrors what the pre-split patch loop computed at the end
/// of the function — final mapfile stats plus whether `sync_all`
/// failed on a regular file (the only kind of fsync error patch ever
/// surfaced; `/dev/null` and pipes always fail `sync_all`, that's not
/// a real error).
pub(super) struct PatchSummary {
pub stats: MapStats,
}
/// Consumer-side of the patch pipeline. Owns the ISO writeback file
/// and the mapfile; publishes a shared snapshot after every record so
/// the producer can read `bytes_good` for stall detection and
/// progress reporting.
pub(super) struct PatchSink {
file: crate::io::WritebackFile,
map: Mapfile,
/// Whether the output is a regular file (so a `sync_all` failure
/// is real). `/dev/null` etc. always fail `sync_all`; ignore those.
is_regular: bool,
/// Snapshot the producer reads. Updated after every successful
/// `record()` call. `Mutex` rather than separate atomics because
/// the producer wants stats + bad_ranges as a coherent pair.
shared: Arc<Mutex<SharedPatchState>>,
}
impl PatchSink {
/// Open `path` as a [`crate::io::WritebackFile`] and pair it with
/// `map` for the consumer. The producer holds onto the returned
/// `Arc<Mutex<SharedPatchState>>` so it can poll mapfile state
/// while the consumer is mutating it.
pub(super) fn new(
path: &std::path::Path,
map: Mapfile,
is_regular: bool,
) -> Result<(Self, Arc<Mutex<SharedPatchState>>)> {
let file = crate::io::WritebackFile::open(path).map_err(|e| Error::IoError { source: e })?;
let shared = Arc::new(Mutex::new(SharedPatchState::from_map(&map)));
let shared_clone = shared.clone();
Ok((
Self {
file,
map,
is_regular,
shared,
},
shared_clone,
))
}
fn republish(&self) {
// Best-effort lock — only the producer reads, only the consumer
// writes; contention is single-acquire so the lock is never
// poisoned in practice. If it ever did get poisoned we'd want
// the underlying error surfaced rather than silently swallowed,
// so we propagate the poison panic. (Same posture as
// `sweep_pipeline.rs` — it never recovers from a poisoned
// mutex either.)
let mut guard = self.shared.lock().expect("PatchSink shared state mutex poisoned");
*guard = SharedPatchState::from_map(&self.map);
}
}
impl Sink<PatchItem> for PatchSink {
type Output = PatchSummary;
fn apply(&mut self, item: PatchItem) -> std::result::Result<Flow, Error> {
match item {
PatchItem::Recovered { pos, buf } => {
let len = buf.len() as u64;
self.file
.seek(SeekFrom::Start(pos))
.map_err(|e| Error::IoError { source: e })?;
self.file
.write_all(&buf)
.map_err(|e| Error::IoError { source: e })?;
self.map
.record(pos, len, SectorStatus::Finished)
.map_err(|e| Error::IoError { source: e })?;
}
PatchItem::Unreadable { pos, len } => {
self.map
.record(pos, len, SectorStatus::Unreadable)
.map_err(|e| Error::IoError { source: e })?;
}
PatchItem::NonTrimmed { pos, len } => {
self.map
.record(pos, len, SectorStatus::NonTrimmed)
.map_err(|e| Error::IoError { source: e })?;
}
}
self.republish();
Ok(Flow::Continue)
}
fn close(mut self) -> std::result::Result<Self::Output, Error> {
// Drain in-flight writeback then issue a full fsync. A failure
// here matters only on regular files — pipes / `/dev/null` etc.
// always fail `sync_all`.
if let Err(e) = self.file.sync_all() {
if self.is_regular {
tracing::warn!(
target: "freemkv::disc",
phase = "patch_sync_failed",
error = %e,
os_error = e.raw_os_error(),
error_kind = ?e.kind(),
"patch: sync_all failed"
);
return Err(Error::IoError { source: e });
}
tracing::debug!(
target: "freemkv::disc",
phase = "patch_sync_skipped",
error = %e,
"patch: sync_all failed for non-regular file; ignoring"
);
}
self.map
.flush()
.map_err(|e| Error::IoError { source: e })?;
// Final republish so anyone reading the shared snapshot after
// `Pipeline::finish` sees the post-flush state. (The producer
// already has its own copy of the final `MapStats` in the
// returned `PatchSummary`, but the snapshot is part of the
// public-ish contract of the consumer: it stays current
// through close.)
self.republish();
Ok(PatchSummary {
stats: self.map.stats(),
})
}
}