Teardown flagged the FUSE worker pool closed but never joined it, so the
process could drop the tokio runtime while a worker was still inside a
job — which is where 2026-09-18's "A Tokio 1.x context was found, but it
is being shutdown" panic came from. `Workers::stop_and_join` now waits
for the jobs in flight, bounded at 10 s so a worker wedged on the network
cannot hold the stop the user asked for; what is still stuck is named in
the log and shutdown continues around it.
The queues stay unbounded — the only caller that could feel backpressure
is fuser's dispatch loop, and blocking that is the stall the pool exists
to avoid — but a lane now logs its depth every 256 jobs, so a queue that
keeps climbing is visible before it is reported as a slow mount.
Documents the hang as B91, with the journal evidence, what made it
undiagnosable, and the candidates the evidence does not yet separate.
A supervisor thread ticks every 15 s and does three things: it WARNs
about a worker that has held one job for two minutes, a control request
that never returned, or a lane that has not completed anything while its
queue is non-empty; it samples resident size and the pending-op count
into the log every five minutes; and it answers the systemd watchdog.
The watchdog ping is deliberately not a heartbeat from a thread that
does nothing else — that would keep the unit "healthy" through exactly
the hang it is meant to catch. It goes out only after a real round trip
over the control socket, so it means "a client can still talk to this
daemon", which is the property that failed on 2026-09-18.
`sd_notify` is spoken directly rather than through libsystemd (READY,
WATCHDOG, STOPPING), including the abstract-namespace address; off
systemd every call is a no-op.
The unit gains `NotifyAccess=all` (rather than `Type=notify`, which
would fail the unit on the start timeout when no login is stored yet),
`WatchdogSec=120`, `Restart=always` — a hang is not a failing exit code
— and `MemoryHigh`/`MemoryMax` as a backstop against the 5.2 G growth
the hung run showed.
On 2026-09-18 the daemon hung: 43 minutes of CPU over 20 hours, 5.2 GiB
resident, silent from 17:10 until a manual restart the next day, and no
panic. Nothing in the journal said what it had been waiting on, and
nothing in the daemon could be asked.
- The worker pool publishes per-thread state: whether the thread is busy,
how long it has been in that state, and the label of the job in hand.
Every FUSE handler that hands work to the pool now names its job
(`read`, `lookup`, `rename`), so a stuck worker is named rather than
merely counted. Queue depths and per-lane completion counts are
mirrored outside the queue lock.
- `Request::Diagnostics` / `pdfs diagnostics` reports those, plus
resident size, the control handler count against its limit, the
in-flight control requests with their ages, and the pending-op count.
- Built to answer while the daemon is wedged: atomics and `try_lock`
only, and the one field that needs the database reads it on a throwaway
thread with a 500 ms deadline.
The trash is a flat list of uids, but a delete has a shape: a folder, and
often items inside it that were trashed separately beforehand. Restoring
one of those in isolation is not what it means — a folder came back
without the contents deleted with it, and a file came back inside a
folder that is still trashed, which is to say nowhere the user can see.
- Schema 29: `trash.parent_uid`, filled by the trash refresh from the
node's parent, plus an index for the walk.
- `expand_restore` takes the connected piece of the trashed tree around
what was asked for (every trashed descendant, every trashed ancestor)
and orders it into waves, shallowest first: the server cannot put a
node back under a parent that is still trashed. A uid the listing does
not know about is restored on its own, because the listing is
materialised in chunks and can be stale.
- `Core::restore` restores wave by wave, still streaming per-node
outcomes, so one node the server refuses leaves the rest restored.
The SDK bump brings the 6cbf2f44 port: permanent API errors are no
longer retried on block transfers, expired block URLs are re-minted on
the whole-file and public-link download paths, and a server-supplied
Retry-After is capped at 120 s.
The "This computer" indexer walks $HOME, and `default_excludes` only kept it
out of the Drive mountpoint and our own state/cache dirs. A home directory
that also holds third-party FUSE mounts (sshfs, rclone) got walked into.
`scan` does ask for `same_file_system(true)`, which looks sufficient and is
not: `ignore` implements it as `is_same_file_system` -> `device_num(path)` ->
`path.metadata()`, run after the directory is queued. Learning that a mount is
foreign requires stat'ing it first, and against a mount whose server is gone
that stat never returns.
The consequence is worse than a stalled scan. The blocked thread sits in
uninterruptible FUSE wait, so it ignores SIGKILL: the daemon can never exit,
systemd's stop times out, and the unreapable process keeps the cache.db lock
the next start needs, turning one dead mount into a permanent restart loop.
No userspace recovery exists once a thread is in that state, so the fix has to
be to never enter it.
Exclude every mountpoint strictly below the scan root, read by path from
/proc/self/mounts (octal-unescaped, so a path with a space still matches).
`filter_entry` runs before the device stat, so the walker never touches them.
Re-read per scan rather than cached at startup: a mount that appeared since
the daemon started is the likeliest to be half-alive.
This does not change what is indexed — a nested mount is a different device,
so it was already going to be skipped. Only the stat in front of that decision
goes away.
Refs: docs/BUGS.md B90
- keyring 3 -> keyring-core 1 + dbus-secret-service-keyring-store 1
(crypto-rust). Same libdbus Secret Service backend; with no target the
store matches items keyring 3 wrote, so stored sessions survive.
- ksni 0.2 -> 0.3 via the blocking API, with assume_sni_available so a
tray that starts before the SNI watcher waits instead of failing. Drops
the clap 2 / atty chain (Dependabot atty advisory).
- gtk4 0.11, libadwaita 0.9, webkit6 0.6, rusqlite 0.40, fuser 0.18,
directories 6, sha1/sha2 0.11, base64 0.23.
- A queued rename that also changes folder now lands as one `move-multiple`
request carrying the target name instead of rename-then-move, so there is
no half-applied state and no second call for stale name requirements to
hit (B46 follow-up). A destination name collision moves under a conflict
name in one step; same-folder renames still use the rename endpoint.
- Shared with me shows the inviter and invite date (GUI subtitle, `pdfs
shared-with-me`), with a warning when the invitation signature does not
verify against the claimed inviter's keys. `DirEntry` gains `shared_by`,
`shared_at` and `shared_by_unverified`, all serde-defaulted.
- Bump `proton-sdk` / `proton-drive-rs` to 0.6.3 (anonymous-node moves,
`TrashTime` trashed detection).
The login backoff had no wake-up path. A unit started at boot sits in
wait_for_session and reaches the ceiling within ~90s; the user then signs
in and `enable --now` is a no-op on an already-active unit, so the daemon
kept sleeping and the mount appeared up to a full interval late.
- service::enable_start now runs `enable` + `restart` instead of
`enable --now`, so the GUI login and the tray's Connect interrupt the
sleep immediately and still start a unit that is down.
- `pdfs login` restarts the unit when it is already active. It stays out
of `enable` on purpose, so a headless setup driving `pdfs daemon` by
hand is left alone.
- Lower the backoff ceiling from 60s to 15s. It is still an 80% cut in
keyring/D-Bus wakeups (28,800/day -> 5,760/day) and it bounds the
worst-case latency for paths that never touch systemd.
Also corrects the flat-3s-poll figure in the doc comment: 86400/3 is
28,800 checks a day, not 48,000.
B69 — spurious `(sync-conflict)` copies came from `(mtime, size)` being used as a
revision identity. The server re-stamps a sealed revision with its own commit
time, so the drain diverted a write into a conflict copy of its own base; and
`plan_file` classified a no-baseline (local, remote) pair as `Conflict`, forking
every pre-existing file on first reconcile.
- `Baseline` carries `revision_id` (from the SDK's new `active_revision_id`),
captured at write-open and preserved across supersede/rebaseline.
`revision_changed` trusts the id and only falls back to `(mtime, size)` for
old sidecars.
- Planner gains `FilePlan::AdoptBaseline`: equal-size no-baseline pairs are
recorded, not conflicted. Divergent sizes stay a real conflict.
- New `sweep.rs` (`pdfs-conflict-sweep`, 5-min cadence) removes conflict copies
proven identical by size + `content_sha1` and surfaces divergent ones once as
an `ActivityKind::Conflict` entry. Never removes what it cannot prove.
B70 — browsers downloading into an on-demand mount had their in-flight temps
sealed as canonical revisions, exiling the finished file to a conflict copy.
`syncignore::is_transient_name` now recognises download/scratch temps, and the
on-demand `release`/`rename` paths hold them (`set_create_hold`) instead of
sealing until the name is finalized.
Workspace deps bumped to proton-sdk/proton-drive-rs 0.2.2 (registry, no patch).
Live validation against a real account still pending for both.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V8P42asymU158CPnhjRwri
- Durable scratch/staging publication and preservation on failure.
- Transactional pending-operation supersession.
- Conflict copies cannot overwrite older copies or lose the source.
- Single-entry wipe protection and incomplete-scan aborts.
- Failed local deletion retains its baseline.
- Downloads are fsynced before baseline settlement.
- Future database schemas are refused.
- Event cursors advance only after successful invalidation and persistence.
- IPC frame, timeout, and concurrency limits.
- Atomic config saves; malformed configs are preserved.
- Sensitive directories require correct ownership, real directories, and 0700.
- Debian package now includes service and tray autostart files.
- Release workflow enforces versions, tests, lint, formatting, and offline acceptance.
- CI now runs offline filesystem acceptance.
- B23: Generate synthetic StagedWrite sidecars in rescue_scratch when sidecar is corrupted
- B26: Add open_count and unlinked state fields to defer node removal and cache eviction on unlinked open files until release
- B27: Discard orphaned write revisions for local UIDs unlinked prior to upload
- B28: Add emergency_evict to ContentCache for LRU blob eviction upon ENOSPC
- B29: Implement FUSE forget handler and forget_lookup in state to prune unreferenced inodes
- B33: Quarantine failing drain ops on HTTP 40x / max retries to unblock offline sync queue
- docs: Track audit items B18-B36 in docs/BUGS.md with fix details and verification results
- Standardize POSIX unlink (EISDIR) and rmdir (ENOTDIR/ENOTEMPTY) checks
- Defer local directory eviction in mode switch until FUSE mount verification succeeds
- Enforce fsync before atomic renames in ContentCache store operations
- Add directory ancestor loop check in rename (EINVAL)
- Implement punch_hole interval tracking in fallocate
- Purge un-created local node write revisions on unlink
- Fix control socket permission race with octal umask 0o77
- Add unit tests for state, DB, and FUSE POSIX logic (208/208 tests passing)
Scripts and health checks had to scrape formatted text, which changes whenever
the display does.
--json is global and applies to status, ls, pins, sync list, devices list,
transfers, activity, and cache inspect. Action commands keep their human
output; a script that needs to know whether one worked has the exit code.
Two decisions worth recording:
The daemon's own reply is serialized rather than a hand-built parallel
structure, so new fields reach the JSON without anyone remembering to add them
-- but the variant tag is stripped first. Serializing CtlResponse directly
yields {"SyncFolders": {"items": [...]}}, which would make every caller write
jq '.SyncFolders.items[]' and couple it to the name of an internal enum
variant. Callers get {"items": [...]}.
A daemon-side error prints its payload and then fails the command. The first
cut exited 0 after printing the error body, which was caught by running
`pdfs --json ls NoSuchFolder`: a script would have read a failed lookup as
success. The body is still emitted so `kind` is available, but the exit code
is what `set -e` and `if pdfs ...` actually test.
status composes its object by hand instead, being the one query that merges
the local keyring with the daemon -- including the logged-in-but-no-daemon
case, which has no daemon reply to pass through at all.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JFzNQyb4TWrCRWcnegqWxY
Verified against a live install: vacuum reported "34.3 MiB freed (170.3 MiB
-> 136.0 MiB), 0 WAL frame(s) checkpointed" while the WAL grew from 67 MiB to
143 MiB. Net disk use barely moved, and the frame count was zero because the
checkpoint ran before the VACUUM, when there was nothing to fold.
Reordering is strictly better, but the commit message should not overstate it:
the live inflation happens because the daemon's other connections hold read
transactions and a TRUNCATE checkpoint waits on readers, so under a busy
daemon this can still move little. journal_size_limit stays the backstop --
the live WAL did fall back to its 64 MiB cap on its own afterwards.
The accompanying test deliberately does not assert on WAL size or frame count.
Measured both ways, a single-connection test reports frames=0 and a 0-byte WAL
under either ordering, because SQLite folds the log itself without a
concurrent reader to block it; an assertion there would pass on the bug and be
decoration. It asserts what is reliably true instead -- that a file-backed
vacuum reclaims space and leaves the database sound -- and the ordering
rationale plus the residual uncertainty are documented on vacuum() itself.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JFzNQyb4TWrCRWcnegqWxY
A corrupt database or an oversized cache previously left the user with no
answer but "delete the state directory and re-sync", which for a large
account means re-downloading everything to fix a problem that may have been
a few stale rows.
pdfs diagnose reports on paths, keyring, daemon, and mountpoint, and exits
non-zero if any check fails so it works in a health check. Every check is
written to run with no daemon, because that is the situation it gets run in:
a stopped daemon is reported as a finding rather than aborting the report.
Writability is tested by writing, since permission bits say nothing about a
full disk or a read-only remount.
pdfs cache inspect reports database size, reclaimable space, per-table row
counts, and cache usage; --deep adds SQLite's integrity check, opt-in because
it reads every page. The report distinguishes "checked and clean" from "not
checked" rather than letting an empty problem list imply a clean bill of
health it never gave.
pdfs cache vacuum checkpoints the WAL and compacts the database.
Deliberately no orphaned-cache-row pruning: ContentCache::reconcile already
rebuilds the index from what is on disk at every open, so a second pass here
would duplicate it against a worse view of the truth. The WAL is likewise
left to journal_size_limit, which already caps it -- a periodic checkpoint
would paper over a problem that no longer exists.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JFzNQyb4TWrCRWcnegqWxY
Mirror folders walked and uploaded every path under their root, so syncing a
project directory pushed node_modules/, target/, and .git/ to Drive -- quota
spent on files nobody wants there, and enough API traffic to invite rate
limiting.
Rules come from a .pdfsignore (or .protonignore) at the folder root, unioned
with an ignore_patterns list in config.json that defaults to the usual build,
VCS, and editor-leftover globs. Both are re-read at the start of every pass,
so edits apply without restarting the daemon.
The filter deliberately sits on the classification union in
classification_order(), not on walk_local alone. do_reconcile classifies
local | remote | baseline, so a file synced before a rule started matching it
is absent from local but present in the other two -- the "local deleted,
remote untouched" arm, which calls trash_nodes. Filtering only the walk would
have meant that adding a line to .pdfsignore silently trashed files on Drive
on the next pass. Ignored baseline rows are left in place so un-ignoring
re-adopts the existing remote file instead of re-uploading it.
guard_local_wipe takes a filtered baseline for the same reason: an ignored
path is absent from the walk by rule rather than by loss, and counting those
as missing would trip the guard on every pass once a rule covered the tree,
wedging that folder's sync permanently.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JFzNQyb4TWrCRWcnegqWxY
nvrl
