| `claimable-trim-miss-test.sh` (f-warn, Zygo's item 1, 2026-10-19) | a trim on a converted legacy raid5 debits every wholly free stripe it takes out: no "above scanned" WARN, and no over-count that the debug kernel's miss counter does not explain | 111 (+ candidate fix) | ~10 min per 300 s iteration | Instrumented #5 build. Fix off, no delay, 2 it: 30 missed debits, 0 WARN. Fix off, 500 ms delay, 3 it: FAIL, 30 WARNs, all explained. Fix on, same delay, 3 it: PASS, 0 WARN. Candidate kernel, 1 it, no delay: PASS (WARN count only). Acked data verifies in every run | 2026-10-19/20 | ehive fwarn #1-#3 (568c45b64012 + instrumentation); candidate zb64 bb9f4c113013 | f1c7f70 | forensics/f-warn-20261019/ |
| `stripe-reclaim-test.sh` with the reclaim debug lines (dynamic debug on / off) | patch 113's "reclaim: relocating|compacting / relocated|compacted block group" lines appear one per reclaim when enabled and not at all when not | 113 | 37 min (paced) / 8 min (unpaced) | PASS: on, misc-next: 19 start + 19 end lines, reclaim_count +19, 0 errors; off (RT_PACE=0): 0 lines, reclaim_count +15. 6.18: on 14 + 14, reclaim_count +14; off 0 lines, +12. (On 6.18 relocation already logs "relocating block group" at info level, so the script's reloc= counts it twice there; verdict unaffected) | 2026-10-21 | misc-next zb64 3e1390a9de03 + 113 + 114 (candidate); 6.18 zb64 68a671e005da + 113 + 114 | 92a2fd2 | forensics/f-oct-20261019/reclaim/seq1, seq618 |
| `stripe-reclaim-now-test.sh` | /sys/fs/btrfs/<uuid>/stripe_reclaim_now: "all" and a named group, the forced kind, the error codes | 114 | 15 min | PASS on both lanes: background reclaim off, nothing moves in 60 s; the least-used group relocated alone (reclaim_count +1); "all" queued 2, trapped 1256 -> 6 MiB; a 79%-used group with trapped stripes compacted; ENOENT, EINVAL (not a number, raid1 metadata group), EROFS, EAGAIN during a paused balance and a running replace; data verifies, no WARN, btrfs check clean. Not shown to fail without the patch (the knob does not exist there: exit 77) | 2026-10-21 | misc-next zb64 3e1390a9de03 + 113 + 114; 6.18 zb64 68a671e005da + 113 + 114 (candidates, promoted 2026-10-21) | 0d0cd25 | forensics/f-oct-20261019/reclaim/seq1/rn.log, seq618/rn.log |
+| `reclaim-futile-test.sh` (RF_FILL_RECLAIM=grow; 3 traced runs per kernel) | the reclaim worker at the fill edge: bounded ENOSPC retries, and a queued (`--enqueue`) resize/replace gets the exclusive operation between groups | reclaimbackoff candidate (2 patches after 114) | 30 min per run | Criterion changed 2026-10-22: "--enqueue wait <= 2 x longest reclaim group (measured per run)"; the old fixed 120 s figure is reported alongside. Lane without the patches (zb64 3e1390a9de03 + 113 + 114): the worker held the operation for a whole sweep (184, 176, 616 s); the first queued resize waited 184, 176, ~655 s (FUTILE, FUTILE, stuck); one-shot resizes 27/40, 28/40, 0/40; 10/10/14 groups per 10 min. Candidate (same + both patches, 12 s gap): every queued resize got in at the first hand-off after it arrived, 0.3-9.4 s after the finish; longest queued waits 82, 11, 23 s, limits 254/54/70 s (all within 120 s): BOUNDED x3; one-shot 12/40, 31/40, 27/40; 13/11/10 groups per 10 min; one ENOSPC backoff (5 errors, reclaim_backoffs +1). Earlier (lane, first shape): 226 reclaim errors in 600 s | 2026-10-22 | misc-next zb64 3e1390a9de03 (+113, 114) vs that + the candidate (tree = refs/cand/reclaimbackoff/mn/zb64) | 9f79e62 | forensics/f-oct-20261019/reclaim/seqTA, seqTV |
Negative controls (2026-10-19): one kernel, 568c45b64012 with patches 67,
99, 109, 110 and 111 reverted and patch 93's commit-before-wait removed