fstools, branch master

block: preen a registered volume before mount on request

2026-07-25T13:59:17Z

blockd gains an optional check_fs attribute on the hotplug registration. When set, the autofs mount path runs the filesystem's preen check (e2fsck -p, fsck.f2fs -a) before mounting, so a volume left dirty by power loss is repaired instead of mounted read-only or refused. The check self-gates: a clean fs is a no-op. Plain fstab mounts keep the global check_fs; the per-mount attribute lets blockd-managed volumes (e.g. uvol) opt in without an fstab section, and does not touch unrelated autofs mounts. blockd hands the attribute down as a check_fs token appended to the mount options of "block autofs add". parse_mount_options() consumes the token like a flag, so it never reaches mount(2), and as a side effect an fstab section can now request the check per mount using 'option options check_fs' as well. Signed-off-by: Daniel Golle

block: use preen mode for the f2fs filesystem check

2026-07-25T13:59:17Z

check_fs is meant to express "repair the volume if it is dirty", which is what the other filesystems get: e2fsck -p leaves a clean volume untouched. fsck.f2fs -f instead forces a full check of the volume on every single mount, which is expensive on large volumes and needlessly delays boot. Use fsck.f2fs -a, which self-gates on the dirty flag just like e2fsck -p. Signed-off-by: Daniel Golle

blockd: do not let a slow startup helper hold back readiness

2026-07-25T13:59:17Z

Readiness is currently announced only once the last coldplug helper has exited. A single helper can take arbitrarily long, most notably when it runs fsck on a large volume, and it must not be interrupted: aborting a repair halfway through leaves the volume in a worse state than not mounting it at all. Until then no mount.ready is sent and the status method keeps reporting false, so every consumer is back to guessing with a timeout of its own. Arm a 10s timer at startup and latch readiness when it expires, leaving the remaining helpers running. The devices they register still emit their own notifications, so a consumer waiting for one specific volume can wait for that instead of for a level which may be minutes away. Signed-off-by: Daniel Golle

blockd: expose storage readiness via a queryable status method

2026-07-25T13:56:46Z

Storage readiness is a level ("is storage available now?"), but blockd only emitted it as an edge: a single mount.ready notification sent once the initial coldplug settles. A subscriber that connects after that notification fired never learns that storage is ready and falls back to a timeout. Track readiness in a blockd_ready latch and expose it via a no-arg "status" method returning { "ready": }, and add the same field to the no-device "info" reply. Consumers can then subscribe and query the current level instead of having to catch the one-shot notification at the right moment. mount.ready stays as a lightweight "something changed, re-query me" hint. Signed-off-by: Daniel Golle

blockd: pass registrant mount options down to block on autofs mount

2026-07-25T13:56:18Z

A runtime volume registered over the block.hotplug ubus call carries no fstab entry, so block had no way to learn how it should be mounted and fell back to read-write. A read-only backing store (e.g. a squashfs or erofs uvol volume, or an ext4 volume registered read-only) then failed to mount with EACCES. Keep the declared options with the device in blockd and hand them to "block autofs add" as a fourth argument when the kernel requests the mount. block applies them to the synthesised runtime mount via parse_mount_options, so "options":"ro" mounts the volume read-only without block or blockd ever inspecting the filesystem type. Signed-off-by: Daniel Golle

blockd: notify when an idle autofs mount is unmounted

2026-07-25T13:56:18Z

autofs_expire() tears down idle mounts (the holder is gone), after which the backing block device has no remaining holder. Emit a mount.umount notification with the device name so subscribers can act on the device becoming free - e.g. uvol's deferred reap of a volume that was still in use when removal was requested. Signed-off-by: Daniel Golle

block: mount ubus-registered devices without an fstab entry

2026-07-25T13:56:18Z

Devices injected into blockd at runtime over ubus (for example by uvol, which derives its volumes from the self-describing LVM/UBI backing store rather than from persistent UCI) have no matching section in /etc/config/fstab. mount_device() looked the device up via find_block() and, finding nothing, fell through to anon-mount or skipped it entirely, so such a device never mounted at its registered target. The autofs trigger only fires for devices blockd has registered for autofs, so its presence is sufficient proof that a runtime registration exists. Synthesise a minimal autofs mount in that case and let the regular autofs path mount the device at /tmp/run/blockd/, which blockd's symlink already exposes at the requested target. This lets a volume be mounted purely over ubus, with no persistent fstab state. Signed-off-by: Daniel Golle

blockd: coldplug block devices

2026-07-25T13:56:18Z

blockd may be started after the hotplug scripts of already attached storage devices have run, leaving those devices unregistered. Call 'block hotplug' for each of them when blockd starts (ie. coldplug). Do not wait for the helpers synchronously: block's autofs and hotplug actions call back into blockd over ubus with a 3s timeout, so blocking the main loop would make every child time out before it can register anything. Reap them via uloop and start the next one from the previous one's callback, so the loop stays alive while at most one helper runs at a time. Serialising them keeps the mount order deterministic where two devices resolve to the same target, and avoids spawning one blkid probe per block device at once on a low-memory system. Announce readiness once the last helper has finished, so the coldplugged device list is complete before subscribers re-register. Signed-off-by: Daniel Golle

blockd: retry interrupted waitpid() in synchronous block() helper

2026-07-25T13:55:08Z

The synchronous variant of block() reads WEXITSTATUS(status) straight after waitpid(), but waitpid() may be interrupted by SIGCHLD from one of the asynchronous hotplug-call children, returning -1 with errno EINTR and leaving status untouched. The garbage status was then decoded as a non-zero exit code (typically 127), so a perfectly successful "block autofs add" was reported as a failure and the autofs request was answered with AUTOFS_IOC_FAIL. The kernel then denied access to a volume that had in fact just been mounted, which surfaced as intermittent first-access failures on freshly registered uvol volumes. Retry waitpid() on EINTR and initialise status so the child's real exit code is used. Signed-off-by: Daniel Golle

blockd: fix bounds of the ubus event name buffer

2026-07-25T13:55:08Z

The third argument of strncat() is the space left in the destination, not the size of the destination. Passing sizeof(evname) - 1 allows up to 15 bytes plus a terminator to be appended behind the "mount." prefix, ie. up to 22 bytes in a 16 byte buffer. No caller passes an action long enough to overflow it today, which is why this has not been noticed, but the bound is wrong. Build the event name with a single snprintf() instead, which cannot write past the end of the buffer regardless of the action. Signed-off-by: Daniel Golle