Checking and repairing filesystems: fsck, xfs_repair and btrfs check on an unmounted filesystem

Goal

Check a Linux filesystem for corruption and repair it only when the tool and mode used are safe for that filesystem's state.

Prerequisites

The target filesystem unmounted (boot into rescue/emergency mode or a live/rescue system first if it is the root filesystem); root access; recent backups if a repair is expected to be destructive.

Steps

1. Confirm the filesystem is unmounted before checking it. e2fsck's manual states plainly that it should not be run on mounted filesystems, and that even where -n makes this technically safe, results printed while a filesystem is mounted are not valid. The generic fsck front-end calls the filesystem-specific checker (fsck.ext4 etc.).
2. For ext2/3/4: run e2fsck -fn /dev/<device> first (forced full check, no changes), then e2fsck -f /dev/<device> to repair interactively, or -y to answer "yes" to every prompt non-interactively — which also accepts destructive fixes.
3. For XFS, generic fsck calls fsck.xfs, which normally just exits 0; use xfs_repair, which refuses a mounted filesystem. First run xfs_repair -n /dev/<device> — "no modify mode": it reports what would be repaired without changing anything and returns 1 if corruption was detected, 0 if not. Without -n it returns 0 when it completes, 1 on a runtime error (re-run it), and 2 when it cannot proceed because the log is dirty. For a dirty log, mount and unmount the filesystem once (on a machine of the same CPU architecture) so the kernel replays the log, then re-run. xfs_repair -L zeroes the log instead: all metadata updates in progress at the crash are lost, which the manual warns may cause significant damage — last resort only, after an image backup.
4. For Btrfs: btrfs check does not modify the device by default (--readonly only reaffirms that); btrfsck is a deprecated alias. Its manual advises against --repair "unless you are advised to do so by a developer or an experienced user," and notes that no checker can fix every kind of corruption. Prefer restoring from a snapshot or backup over --repair where one exists; on a mounted filesystem, btrfs scrub start <mountpoint> verifies checksums online, which is not a structural check.
5. To force a check of the root filesystem during one boot, add fsck.mode=force at the boot menu (kept in the configuration it forces a check at every boot; auto is the default, skip disables checks); fsck.repair= (preen, yes, no) controls how aggressively systemd-fsck answers repair prompts during that automatic check.
6. Interpret the generic fsck exit status as a bitwise sum: 0 no errors, 1 errors corrected, 2 system should be rebooted, 4 errors left uncorrected, with higher bits for operational failures; codes of several filesystems checked in one run are OR'd together, so check one filesystem per invocation when a script needs a precise result.

Expected result

A clean pass reports no corruption (exit 0, or xfs_repair -n exit 0); a repair pass leaves the filesystem mountable and the tool's own summary states what was fixed.

Limits and test basis

XFS and Btrfs define their own exit codes and options — do not assume they match fsck's. btrfs check --repair does not yet work together with --mode=lowmem. Back up (or snapshot) before any non-dry-run repair; a checker can make a damaged filesystem worse.