Skip to content

19-livelock-fix.md

Latest commit

 

History

History
102 lines (81 loc) · 5.07 KB

19-livelock-fix.md

File metadata and controls

102 lines (81 loc) · 5.07 KB
Raw

LIVELOCK FIX — Phase 1: the three defects (+ signal mode)

Applied on top of Phase 0b (4th fix + audit). Machine exclusive for the gate runs (load ~0.1, own runs only).

Defect 1 — lo_declared_dist() origin is now signal-mode-aware

The user's original Defect 1 (unconditional d starts at 0) was wrong for the post-consumption path and is not what was implemented. The correct origin depends on when pager_notify_access() fires relative to the read it announces:

mode when notify fires d starts at dist(seq[pos])
post (policy_set_signal_mode(0), default) after seq[pos] fully read 1 seq_len (furthest)
pre (policy_set_signal_mode(1)) before seq[pos] read 0 0 (imminent)

policy.c:lo_declared_dist() now scans for k in 0..seq_len-1: d = d0 + k with d0 = signal_mode ? 0 : 1. d0 == 1 is byte-for-byte the old for d = 1..seq_len loop — post mode is a no-op refactor. Chunks absent from the declared sequence still fall through to INT64_MAX in both modes. --protect-current is unchanged and orthogonal (forces seq[pos] / seq[pos-1] to 0 regardless of mode).

Wiring:

  • replay_main.c: policy_set_signal_mode(consumption_signal_all ? 0 : 1) — all-threads (the default) → post; tid0 → pre. The serial replay_cyclic path ignores consumption_signal_all and thus stays post under the default, a deliberate baseline-continuity choice (docs/design-history.md).
  • residctl_llama.c: policy_set_signal_mode(1) — the real model, because Defect 2 moves the notify to the pre-compute callback pass.

signal_mode=pre|post is now emitted in ARM_CSV and RESIDCTL_STATS.

Defect 2 — wp2_gen.cpp:eval_cb() acts on the pre-compute pass

if (ask) return true; → if (!ask) return true;. The eval callback fires ask=true before a node computes and ask=false after; the body (which fires residctl_llama_notify_layer / notify_role) now runs on the pre pass, so the consumption signal precedes the weight read. Combined with the 4th fix, token_embd too is signalled pre-read. Token-identity is checked in Phase 3's correctness gate.

Defect 3 — declined-prefetch backoff

chunk_t.decline_until_ns (new). prefetch_pool.c:do_one_prefetch() sets it to now + 100 ms on the ensure_budget_prefetch() != 0 path (covers both stat_prefetch_declined and stat_prefetch_infeasible — the spec equates "decline" with "returns −1"). prefetch_pool_top_up() skips any candidate whose decline_until_ns is still in the future, next to the existing pf_tracked / CHUNK_ABSENT checks. No growing backoff, no counter, no new structure. Expectation 3 in Phase 3 is stated in terms of stat_prefetch_declined alone; the backoff also suppresses re-tries after a true infeasible, so attribute the drop to both mechanisms.

Unit tests (test_policy.c) — PASS

declared next-use vs naive Belady scan (post, protect-on): agree (all positions, all chunks)
declared next-use vs naive Belady scan (pre, protect-off):  agree (all positions, all chunks)
layer_order_declared post protect-off:      ok
layer_order_declared pre protect-off:       ok
layer_order_declared protect-on both modes: ok
PASS
  • post + protect-off (unchanged): dist(seq[pos]) == seq_len, dist(seq[pos-1]) == seq_len-1, dist(seq[pos+1]) == 1.
  • pre + protect-off (new): dist(seq[pos]) == 0, dist(seq[pos+1]) == 1, dist(seq[pos-1]) == seq_len-1, dist(chunk absent from seq) == INT64_MAX, select_victim picks the furthest.
  • protect-on, both modes (new): seq[pos] and seq[pos-1] both 0; absent chunk still INT64_MAX.
  • Belady cross-check, pre mode (new): independent naive scan with k >= 0 agrees at every position, every chunk.
  • The existing post-mode cross-check was insensitive to the d0 change — protect-on short-circuits seq[pos] / seq[pos-1] to 0, and those are the only inputs where starting at 0 vs 1 differs. It passes byte-identical; the new pre-mode cross-check is what actually exercises the corrected origin.

Regression gates — PASS

gate result
WP1 §1.2 (run_wp1_gate.sh, layer_order_learned) absent_handled=57 evictions=49 pager_bytes_fetched=7650410496, all 3 reps — exact, GATE PASS
T-1..T-5 (run_correctness_harness.sh) RESULT: PASS
T-6, T-7 (run_t6_t7.sh) mismatches=0, all storm threads joined, dedup_fetching=16161 > 0 — RESULT: PASS
test_prefetch, test_eviction PASS (Defect 3 path)
Phase 0 diagnostic (Defect 2 pre-compute timing) 320 notifies / 8 tokens, all 40 declared chunks in declared order, cursor 7 clean wraps / 0 jumps, audit OK — GATE PASS

(T-1..T-7 run with reconcile_interval = 1 compiled into the test binaries = --eager-reconcile.)

Files touched

src/policy.{c,h} (Defect 1 + signal mode), src/wp2_gen.cpp (Defect 2), src/region.h + src/prefetch_pool.c (Defect 3), src/replay_main.c + src/residctl_llama.c (signal-mode wiring + signal_mode in the stats lines), src/test_policy.c (both-mode tests), docs/design-history.md (new).