repair). Side findings: scrub on nodatasum data recomputes parity
from the stale strip (both cases); raid6 RMW recovery does not try Q
when P is silently stale. Evidence:
- raid56-wh-harness/forensics/f-ioerr-20261020/ (README.md files).
\ No newline at end of file
+ raid56-wh-harness/forensics/f-ioerr-20261020/ (README.md files).7. NEW (f-oct, 2026-10-21), btrfs-progs `--enqueue` and a briefly freed
+ exclusive operation. `check_running_fs_exclop()` (common/utils.c) reads
+ sysfs `exclusive_operation` before any exclusive ioctl. With
+ `--enqueue` it waits for the notification, but (a) after the notify it
+ reopens the file and sleeps in a second `select()` for
+ `(remaining % 10) + 1` s "to randomize the order" before it re-reads
+ the operation, so a holder that frees the operation for less than ~10
+ s (the reclaim worker between groups) usually retakes it first; and (b)
+ it closes and reopens the file every loop, and a new open takes the
+ node's current event count, so a notify that lands between a select()
+ returning and the reopen is lost until the next notify or the 60 s
+ timeout. The kernel side works around (a) with a 12 s gap between
+ reclaimed groups (reclaimbackoff candidate). Sketch of a progs fix:
+ keep one fd open across the loop (kernfs poll compares the open file's
+ event count with the node's, so no notify is lost), re-read the
+ operation immediately after each notify and go straight to the ioctl
+ when it is free, and randomize only when the ioctl itself loses the
+ race (`BTRFS_ERROR_DEV_EXCL_RUN_IN_PROGRESS`). No progs patch written;
+ want one?