scrub again. The other arms (W, R, D, T) test the raid56-fixes stats
candidate; see that topic.
+## sb1104 (2026-11-04/05): the 6.18 topic now starts with Filipe's three scrub fixes
+
+6.18 only; the series sent upstream is unchanged (misc-next has the three in its base). Zygo 11-04:
+backport Filipe Manana's scrub fixes, in btrfs/for-next and not stable-tagged:
+f64d1d4ceeb7 "fix lost wakeup and hang when cancelling a scrub that is finishing", cc1369667c11 "fix race
+leading to use-after-free in the scrub progress ioctl", ab268571b210 "fix race between scrub and scrub progress
+ioctl". Clean cherry-picks with `(cherry picked from btrfs for-next commit ...)`; they sit after the older for-next
+backport already at the head of 6.18 read-repair-fixes (Qu's repair-bio commit keeps its SHA). raid56-scrub-replace,
+stacked on read-repair-fixes, was replayed: every replayed commit has the same content, message and identity. The
+triage-time scrub.c "conflict" was the third fix picked without the second; in order they are clean everywhere, so
+no octopus resolution. 6.18 read-repair-fixes da2af9dca8a3, raid56-scrub-replace 1e0e948ea5a8, testing
+a34a4f1e68a6, zb64 6b4c738fd38c.
+
+| test | kernel | result |
+|---|---|---|
+| per-commit buildcheck (read-repair-fixes 8, raid56-scrub-replace 13) | bhive | 0 failed, 0 new warnings, tip links |
+| scrubrace progresspin (progress ioctl vs scrub exit; pinned, preempt=full; KASAN) | 6.18 #316 (lane before) | KASAN use-after-free in btrfs_scrub_progress (memcpy of 120 bytes from the freed ctx), 1 in 900 s |
+| same | 6.18 #317 (this zb64, KASAN + lockdep + ASSERT) | 0 in 2700 s (2.5x the scrubs of #316): consistent with the fix, not proof |
+| lost wakeup, stats tear | - | BLIND: btrfs_scrub_cancel_dev() is reachable only from device remove, whose shrink commit waits for a pauseless scrub to end; no KCSAN for the stats copy |
+| rb1028 regression (minus qu-p1), replace-discard, replace-status, scrub-repair-werr, scrub-blame, raid56-devstats P, nodatasum-scrub-repair raid6 W1/C, replace-resume-gap, replace-cancel-persist, scrub-uncorrectable-rc, replace-uncorrectable, failing-member-scrub | #317 | all as on the lane before (ck1104 / rb1103 / the rows above); scrub-fserror-stop: replace PASS, scrub arm INVALID (ended_at_abort: the scrub had already stopped when the test looked, every measurement in the PASS range) |
+
+No kernel report (KASAN/WARN/BUG/lockdep/hung task), debug_locks 1, taint 0. Evidence: raid56-wh-harness
+`forensics/sb1104-618/`; scripts vm-scripts `sb1104/`.
+
## Not covered / open
- raid6 with two bad devices plus a replace (raid5-2dev-replace is raid5
lookup can itself fail after an abort); scrub csum lookups (commit root,
consistent); dev-replace copy (continues on purpose, untouched).
+## sb1104 (2026-11-04/05): the 6.18 topic now starts with Filipe's three scrub fixes
+
+6.18 only; the series sent upstream is unchanged (misc-next has the three in its base). Zygo 11-04:
+backport Filipe Manana's scrub fixes, in btrfs/for-next and not stable-tagged:
+f64d1d4ceeb7 "fix lost wakeup and hang when cancelling a scrub that is finishing", cc1369667c11 "fix race
+leading to use-after-free in the scrub progress ioctl", ab268571b210 "fix race between scrub and scrub progress
+ioctl". Clean cherry-picks with `(cherry picked from btrfs for-next commit ...)`; they sit after the older for-next
+backport already at the head of 6.18 read-repair-fixes (Qu's repair-bio commit keeps its SHA). raid56-scrub-replace,
+stacked on read-repair-fixes, was replayed: every replayed commit has the same content, message and identity. The
+triage-time scrub.c "conflict" was the third fix picked without the second; in order they are clean everywhere, so
+no octopus resolution. 6.18 read-repair-fixes da2af9dca8a3, raid56-scrub-replace 1e0e948ea5a8, testing
+a34a4f1e68a6, zb64 6b4c738fd38c.
+
+| test | kernel | result |
+|---|---|---|
+| per-commit buildcheck (read-repair-fixes 8, raid56-scrub-replace 13) | bhive | 0 failed, 0 new warnings, tip links |
+| scrubrace progresspin (progress ioctl vs scrub exit; pinned, preempt=full; KASAN) | 6.18 #316 (lane before) | KASAN use-after-free in btrfs_scrub_progress (memcpy of 120 bytes from the freed ctx), 1 in 900 s |
+| same | 6.18 #317 (this zb64, KASAN + lockdep + ASSERT) | 0 in 2700 s (2.5x the scrubs of #316): consistent with the fix, not proof |
+| lost wakeup, stats tear | - | BLIND: btrfs_scrub_cancel_dev() is reachable only from device remove, whose shrink commit waits for a pauseless scrub to end; no KCSAN for the stats copy |
+| rb1028 regression (minus qu-p1), replace-discard, replace-status, scrub-repair-werr, scrub-blame, raid56-devstats P, nodatasum-scrub-repair raid6 W1/C, replace-resume-gap, replace-cancel-persist, scrub-uncorrectable-rc, replace-uncorrectable, failing-member-scrub | #317 | all as on the lane before (ck1104 / rb1103 / the rows above); scrub-fserror-stop: replace PASS, scrub arm INVALID (ended_at_abort: the scrub had already stopped when the test looked, every measurement in the PASS range) |
+
+No kernel report (KASAN/WARN/BUG/lockdep/hung task), debug_locks 1, taint 0. Evidence: raid56-wh-harness
+`forensics/sb1104-618/`; scripts vm-scripts `sb1104/`.
+
## Open
- Data writeback after an abort (Zygo's call): pages dirtied before an