Add a third breadth-first tier (PATCH_TIERS 2->3) that runs the marginal
specialists on the hardened residual tiers 0-1 leave: SlowSpin (Linear
fwd+rev @ min), FuaRetry (Linear fwd+rev+Bisect @ FUA), SlowFua (Linear @
min+FUA), CachePrime, Oscillate (@ max and @ min), SpeedSweep. Every read is
a wedge-safe read_span, so they inherit wedge-abort / unproductive-yield /
deadline for free. All are new configs, so the EWMA scorecard calibrates each
once then ranks by decayed rate — a specialist that doesn't fit self-
deprioritises. Tiers 0-1 (fast scouts, slow-deep) are unchanged; this is
purely additive. Also switch the scorecard log sort to sort_by_key.
cargo test -p libfreemkv green (2200 passed).
CachePrime reads the good run immediately preceding a residual island to lock
the servo/PLL, then reads the island warm (boundary sectors the drive can't
cold-seek). Oscillate reads each residual sector by alternating approach —
forward-into then reverse-into — for direction-dependent tracking. Both go
through the wedge-safe read_span (primes included). FakeDisc models a
direction-dependent sector and a servo-primed boundary sector; fixtures prove
a forward/cold Linear misses each while Oscillate/CachePrime recover them.
FuaRetry (Linear fwd+rev+Bisect at {max,fua,deep}) and SlowFua (Linear at
{min,fua,deep}) are Linear/Bisect at FUA params, no new struct. FakeDisc now
models a stochastic sector that lands only on the Nth physical (FUA) read (a
cached re-read keeps missing) and a hardest sector needing BOTH min speed and
FUA. Fixtures prove cached reads keep missing while the FUA group lands the
stochastic sector, and that neither lever alone recovers the hardest sector —
only the min+FUA combination does.
SlowSpin is Linear pinned to min speed (no new struct). SpeedSweep is a new
per-sector handler that tries Max->Min until one reads (speed resonance).
FakeDisc now models a sector that reads ONLY at min speed; fixtures prove a
max-speed Linear misses it while SlowSpin and SpeedSweep recover it (and that
SpeedSweep tries fast-then-slow).
Grade handlers by an exponentially-decayed recovery rate (bytes/second,
alpha=0.5) instead of a cumulative rate. The residual hardens mid-pass, so
the best technique changes: cumulative froze the early winner in the lead
forever; the EWMA re-prices continuously — a handler that stops earning
decays down, a late-starting specialist climbs. Keeps rank()'s untried->top
one-shot calibration and attempted-but-zero-time->bottom.
New flip fixture proves a handler that recovers a lot early then nothing
loses its lead to one that starts recovering later.
Add ReadParams { speed: SpeedPref, fua: bool, timeout: TimeoutPref } and
thread it through read_span so every wedge-safe handler read can request a
spindle speed (SET CD SPEED issued only on change, restored to max when the
handler exits), set the READ(10) FUA bit, and pick the 10s vs 60s timeout.
- SectorSource gains read_sectors_fua (default ignores fua); Drive sets the
CDB bit, DecryptingSectorSource threads fua to its inner read.
- recovery_read gains a fua param.
- Linear becomes { direction, params }; Bisect/Jump take params. Existing
tier-0/1 instances keep identical behavior (max speed, no FUA, fast/deep).
- Scoreboard keys on the full-config String name (linear:fwd:max:fast, ...).
- FakeDisc observes speed + FUA + approach so specialist techniques are
provably exercised in later commits.
cargo test -p libfreemkv green (2193 passed).
- Live located drilldown (sweep + patch progress snapshots) now excludes
NonTried, matching the one-shot path. Including the unread remainder made
main_at_risk_ms show ~full-movie at sweep start and melt to 0 as it
progressed — unread is unknown, not damage.
- Wedge abort now requires the failure to be FAST (< WEDGE_FASTFAIL_MS,
500ms) as well as wedge-family sense: a real uncorrectable sector on
Hardware-error media spends ECC-recovery time before failing, so it no
longer false-trips the fast-fail wedge abort. New regression test.
- Removed the always-zero blocks_attempted/read_ok/read_failed/
unreadable_count from PatchOutcome + PatchLoopState (dead residue from
the old grind loop; the HandlerScoreboard supersedes them) so the
patch.done log no longer emits misleading zeros.
Handlers (section_recover.rs):
- Bisect expand loops now honor ctx.halted() (were deadline-only, so a
Stop could hang up to 60s vacuuming a readable island).
- read_span: explicit Transport arm so a bus-abort read isn't counted as
unproductive grinding; debug_assert the sector-aligned span invariant.
- Scoreboard rank: an attempted-but-zero-time handler (e.g. returned Halted
on its first check) now ranks BOTTOM, not top — it no longer crowds out
proven performers.
- Document the wedge tier-size coupling + new regression test that a
2-handler (tier-1) chain still catches a wedge via cross-section streak.
Pass-N engine (patch.rs):
- Rebuild PatchOutcome stats AFTER the post-read re-verify downgrade flush
(was snapshotting before it, over-reporting bytes_good / recovered and
risking a 'perfect rip' verdict on an imperfect one).
- Progress 'recovered' composes the still-bad set to MATCH work_total
(subtract NonTried, add Unreadable) so the bar can't pin at 0 on a
partially-swept disc or run backward on the Unreadable→NonTrimmed relabel.
- Remove dead work_done field; rewrite the stale 'adaptive batching' comment
to describe the handler chain and mark block_sectors/full_recovery as
informational-only.
Sweep (disc/mod.rs):
- Saturating arithmetic at the damage-jump position math (honor the
read_error side's documented defence-in-depth guarantee).
Deferred (noted, need focused passes): fast_capture re-introduction,
Pass-1 halt-misclassified-as-jump, bytes_good display inflation, the
always-zero blocks_* telemetry, Pass-1 jump-on-first-error policy.
The 2026-07-01 overnight rip ground a wedged BU40N for 28 minutes at
0 B/s. Root cause: the handler chain's read_span split errors only into
Transport (bus gone) vs Bad (everything else), so the drive's fast-fail
wedge sense (ILLEGAL REQUEST 0x05/0x24 — it rejects every CDB in <100ms
without attempting recovery) was treated as an ordinary bad sector. The
chain hopped to the next section forever, and the existing wedge detector
in read_error.rs never saw it because the chain reads through
recovery_read directly.
Add a pass-level wedge streak: read_span counts consecutive wedge-family
(Hardware/IllegalRequest) senses; at WEDGE_ABORT_STREAK (16) it escalates
the read to Transport, which every handler already propagates as
TransportFault — aborting the whole pass and setting wedged_exit so the
caller spin-cycles. The streak is carried across sections via PatchCtx
(seeded into and read back from each per-section HandlerCtx) so a wedge is
caught even when every bad sub-range is smaller than the streak. Any Good
or non-wedge (medium-error) read resets it, so scattered bad sectors on
real media never trip it. New fixture test: a wholly-wedged 1000-sector
section aborts in <100 reads, not 1000.
The expand loops broke out of the batch sweep on the first failing read,
overshooting the good/dead boundary by up to a batch and leaving the
readable sectors adjacent to a dead edge for the re-bisect to re-pin. On
top of that, early-yield counted those boundary-probe failures as a
stall, so the forward expand quit early AND the backward expand inherited
a poisoned unproductive counter and yielded instantly.
Now on a failed expand batch we halve the step (down to a single sector)
to recover right up to the dead edge in ~log2(batch) reads, drive the
expand loops off the deadline only (they self-terminate via halving), and
clear the unproductive streak once an island is located. patch_recovers_
good_middle_of_a_bad_range now recovers 50/50 good-middle sectors.
Scorecard: grade each handler by recovery rate (MB/s) per rip, order
best-first on later sections, log the ranking at pass end. Untried
handlers rank top so each is calibrated once before ranking narrows to
the winners. Ephemeral (reset per pass), no persistence.
Tier 0 scouts are now [Bisect, Jump, Linear-fast x2], scorecard-ordered.
Bisect leads: probing the MIDDLE of a range lands on a readable island in
one read where a linear scan grinds the dead front to reach it. Jump now
jumps to the middle of the REMAINING span (proportional) instead of a
fixed 8 MiB that leapt clean over small ranges and missed their readable
middles. Tier 1 is slow deep-recovery on the residue.
Scorecard (ephemeral, reset each pass, no persistence): grades every
handler by recovery rate (bytes/sec). run_handlers orders handlers
best-first by that rate; an untried handler ranks top so each is
calibrated once, then the ranking narrows to the winners. Logged at pass
end (phase=scorecard) so the operator sees which handler is pulling the
weight on this drive/disc and which is a dud.
Tier 0 scouts are now [Bisect, Jump, Linear-fast x2], scorecard-ordered.
Bisect leads because probing a range's MIDDLE lands on a readable island
in one read, where Jump (linear from the front, big skip) can grind the
dead front or overshoot a small range entirely. The scorecard confirms
or overturns that order with real per-disc data.
Each dead probe read costs the drive's full ~10s timeout, so a large dead
region took a dozen escalating probes. Starting the jump at 8 MiB (vs
1 MiB) clears it in a handful; a skipped span stays bad for Bisect to
reclaim readable islands, so an over-jump loses nothing.
Tier 0 is now a single fast Jump scout: it streams the big readable
ranges back and skips dead runs in seconds, so the pass reaches every
section fast and converges to the small genuine-dead residue instead of
grinding three handlers x 60s on each dead fragment. Tier 1 (fast
mop-up + slow deep reads + Bisect) works only that residue.
Bisect now expands: on a good probe it reads outward forward and
backward in full batches until a read fails, recovering the whole
readable island in large reads; the two failing ends become smaller bad
sub-ranges it bisects again. One huge bad range becomes many precisely
located small dead clusters.
Progress heartbeat: HandlerCtx gains a throttled tick (250ms) called
from every read, pushing a fresh snapshot to the reporter DURING a
handler. The bar and speed now move continuously as recovery happens
instead of jumping once per section (the reason speed read 0 B/s and the
% looked frozen between range boundaries).
Jump: on sustained batch failures skip ahead an escalating distance
(1 MiB doubling to 256 MiB) to find where readable data resumes, leaving
the skipped span for Bisect to pin — mirrors the Pass-1 damage-jump. It
leads the fast tier so a large dead run is skipped in seconds instead of
the linear sweeps grinding every dead batch (10 s each) first; on a
readable range it just streams it back. Recovers readable data buried
behind a big dead front (the 192 MB Dune range).
Progress %: report bytes RECOVERED (initial-bad minus still-pending)
instead of a per-range counter that only advanced on the final tier — so
the bar reflects the readable bulk recovered during tier 0 the instant it
lands, matching the 'MB remaining' number.
Pass-N recovery is now a chain of time-bounded recovery handlers instead
of one monolithic per-range loop that could grind the front of a bad
range for up to 30 min, wedge the drive, and abort the whole pass.
A bad range is a SubRanges set; recovery is an ordered list of
SectionHandlers (Linear{reverse,fast} covering back/forward x fast/slow,
and Bisect). A coordinator runs each handler with a hard per-handler
deadline: a handler recovers what it can (removing it from the still-bad
set) and hands the rest to the next handler; whatever is still bad after
the chain becomes NonTrimmed residue and we move on to the next range.
Guarantees, now structural rather than bolted-on:
- never hangs: every handler is deadline-bounded; the loop always drains
to recovered-or-residue.
- always moves on: a range that cannot be finished leaves residue and
advances; only a genuine transport fault or user halt ends the pass.
- extensible: a new recovery idea is one SectionHandler impl added to the
chain; a proven-ineffective one is removed. The engine never changes.
Removes ~1.9k lines of the old inner loop (watchdogs, skip escalation,
NOT_READY grind, wedge counters) and their tests. fast_capture is now
inert (the chain supersedes it); breadth-first ordering becomes a future
scheduler concern. New module: disc/section_recover.rs (8 fixture tests,
injectable clock — bounded/never-hang proven without touching a drive).
Two A/B tests updated to the chain's strictly-better recovery counts.
Matthew Jackson