nvrl
a9cb389f04
Issue #40: after a password reset the daemon failed every mount with "failed to decrypt: missing key", and the app waited on "Connecting…" because control.sock was never created. proton-drive-rs 0.9.1 now creates a new My Files volume when no active regular volume is left, as the web client does. - BUGS.md: B210 with cause, fix and the SDK tests that cover it - TROUBLESHOOTING.md: what the log line means and how to restore the files of the locked volume from the web
442 KiB
Bug ledger
Every defect and side finding tracked for this client: how it was found, what caused it, how it was fixed, and how the fix was checked. The index sorts the entries by status.
To report a problem, open an issue and include what TROUBLESHOOTING.md lists. An issue that turns out to be a defect gets an entry here.
Numbers
Code comments, tests and acceptance cases refer to entries by number, for example
docs/BUGS.md B123 or the case regression B123. A number is therefore permanent: an entry is
never renumbered or removed, and a number is never reused. B76 and B77 were never assigned.
Findings about the acceptance runner itself are numbered T<n>.
Statuses
| Status | Meaning |
|---|---|
| Open | Not fixed |
| Partly fixed | Part of it is fixed; the entry says what is left |
| Fixed (unverified) | The change is in and has unit or simulation coverage, but nobody has driven the real flow against a live account since. It stays so until someone does |
| Fixed | Fixed and checked: on a live account, by the acceptance suite, or by the test the entry names. A date says when |
| Won't fix | Understood and accepted on purpose; the entry says why |
| Not a bug | The behaviour is correct or the report was mistaken. Kept as a record |
| Closed | No longer applies |
Older entries use other words for the same states, such as "fixed in code", "fixed in-tree" or "live-validated". The index sorts each entry into one of the states above; the entry keeps its original wording.
Adding an entry
- Take the next free number and add the entry at the top of Entries.
- Record how the bug was found, not only what it is. The reproduction is the expensive part.
- Name the test that covers the fix: a unit test, a simulation seed, or an acceptance case named
regression B<n>(see TESTING.md). - Add the entry to the index. When its status changes, update the status line and move it in the index in the same change.
### B196 — What the user sees, in one line
**Status:** Open
**Found:** 2026-10-06, how it was found
**Where:** `crates/pdfs-fuse/src/drain.rs`, `Core::function`
**Repro:** The steps, and what happens.
**Cause:** Why it happens.
**Fix:** What changed, and why that is the right layer.
**Verified:** The test, run or check that shows it, with its date.
Index
Open
| Entry | Title | What is left |
|---|---|---|
| B134 | A file deleted online while the drain checks its write is kept as a conflict copy | Only with local_first off |
| B91 | The daemon hung overnight, and nothing could say what it was waiting on | The cause is not diagnosed; the daemon now reports where it is stuck and is restarted |
| B67 | Fresh-state restore can silently select a new device (HIGH-22) | An ambiguous device after a loss of local state is not asked about |
| B66 | Local-only pending data is not protected at shutdown (HIGH-21) | No shutdown protection and no pdfs sync flush |
| B65 | Release artifacts lack supply-chain verification (MED-11) | No SBOM, provenance or checksums |
| B15 | An empty-but-listed folder, and duplicated folder uids |
Cause not attributed |
| B10 | GLib critical when the launcher closes over an in-flight open |
Partly fixed
| Entry | Title | What is left |
|---|---|---|
| B199 | A stopped share stays on the Shared by me page for minutes | The share itself is not deleted; the SDK has no call for it |
| B88 | pdfs pin, rename, move, delete and sharing still cannot name a path under a secondary mount |
pin, unpin, opening a file, thumbnails, uploads, sharing and versions |
| B72 | pdfs trash never returns |
The slowness of the first refresh is unmeasured |
| B71 | Conflict sweep is not production-ready (auto-trash loop, CRIT-13) | Items 5 to 10 of the entry |
| B68 | Profile backup cannot be written to the device root | Backup health is not shown in the app; the restore was not re-run |
| B64 | Stable publishing lacks data-safety and recovery gates (CRIT-12) | Recovery approval before a release |
| B63 | Release version identity is inconsistent and unenforced (HIGH-20) | Protocol version identity |
| B57 | Event cursor can advance past unapplied state (HIGH-14) | Database and apply fault tests |
| B56 | Database schema version parsing fails open (HIGH-13) | Malformed schema versions |
| B12 | Cold enumeration is slow per entry, and goes superlinear past ~500 | The attribute-invalidation follow-up is unverified |
| B8 | No way to complete a CAPTCHA, so a gated sign-in is unrecoverable | Not tested against a real verification request |
| B2 | The reported .mkv went to trash, the repro vanished entirely | Where the trash came from is still unattributed |
Fixed, not verified live
The change is in and has unit or simulation coverage. Nobody has driven the real flow against a live account since.
| Entry | Title |
|---|---|
| B210 | After a password reset, the mount retries "missing key" forever |
| B209 | A blocked sign-in says the app version is no longer supported |
| B208 | A file landing while its folder is pinned is kept but not available offline |
| B207 | A move between locations goes ahead under a file opened just after its create landed |
| B206 | An upgrade leaves the old daemon running, and nothing says so |
| B205 | A photo upload the SDK refuses says "You're offline" |
| B204 | A write over a partial edit whose base was never read fails with EIO |
| B203 | pdfs unlock leaves a running daemon a spent refresh token |
| B202 | A refused session shows as offline forever, and nothing says to sign in again |
| B201 | After signing in again, the old daemon puts its spent tokens back |
| B200 | Offline with changes queued, the daemon keeps several cores busy |
| B198 | Switching to online-only deletes local files that never reached Drive |
| B197 | Switching to online-only trashes synced files that became ignored |
| B196 | One unreadable folder stops the whole synced folder from syncing |
| B195 | The app crashes when Proton asks for a CAPTCHA on secureblue |
| B193 | Deleting a file whose create failed can trash a file of the user's on Drive |
| B191 | A file deleted after its create failed and was written again can stay on Drive |
| B190 | A file saved just before the daemon stops can keep its old content on Drive |
| B189 | A change that failed on a short outage waits out its retry long after the link is back |
| B188 | A restart can leave a queued write retrying forever on a file deleted while open |
| B187 | An uploaded file can leave a copy of itself in the cache until the next start |
| B186 | A file deleted offline while open fails a partial write with "Input/output error" |
| B185 | A new file written again just as it starts uploading waits out a retry |
| B184 | A file renamed after its create lost its answer stays on Drive under its old name |
| B183 | A file deleted after a restart stays on Drive when its create lost its answer |
| B182 | A write after a restart over a create never read back lands as a conflict copy |
| B179 | A file read just as its create lands reads empty |
| B178 | Trashing a folder while sync is paused keeps a create in it as a trash |
| B177 | A folder listed just as its create lands answers "Input/output error" |
| B176 | A write closed just as its file's create lands is dropped |
| B175 | A file replaced by a rename just as its create lands stays on Drive |
| B174 | A file renamed just as its create lands is refused with "Permission denied" |
| B173 | A file deleted just as its create lands stays on Drive |
| B172 | A file renamed while its upload is read back shows its old name |
| B171 | A file deleted while its upload fails stays on Drive for the upload's backoff |
| B170 | A write opened as an upload lands fails the next write with EIO |
| B165 | A stop waits on a stuck thread until systemd kills the daemon |
| B159 | A write right after a restart fails with "Resource temporarily unavailable" |
| B158 | A file renamed twice in a row ends up under the first name, again |
| B157 | A second write to a file whose upload was not read back fails with EIO |
| B156 | A file written again after its create lost its answer lands as a conflict copy |
| B155 | A write whose upload lost its answer lands as a conflict copy of itself |
| B154 | A file made under the name of a create on the wire lands as a conflict copy |
| B153 | An op waiting for a name sat out its backoff after the name was free |
| B152 | A rename made as its create landed was undone in the tree |
| B151 | A file deleted after its create lost its answer stays on Drive |
| B150 | A conflict copy whose folder is gone retries forever |
| B149 | pdfs rename or pdfs move right after a rename through the mount is undone |
| B148 | A folder pinned before it has reached Drive loses its pin |
| B147 | A new file made under a name just renamed away can share the first file's node |
| B146 | A file renamed just as its upload lands can go back to its old name |
| B145 | Renaming a just-made download to its final name can fail with "Device or resource busy" |
| B144 | A rename can land under a conflict name although its target was freed |
| B143 | A write to a file whose create landed as the link went down lands as a conflict copy |
| B142 | Renaming a new file as its upload lands can fail with "Device or resource busy" |
| B141 | A file renamed twice in a row can end up under the first name |
| B140 | Files in a folder that lands as the daemon stops are gone after a restart |
| B139 | A file closed while its first upload lands keeps its last write in staging |
| B138 | A file created where another is leaving lands as a conflict copy |
| B137 | A queued change can be deleted by an older one landing |
| B136 | A file renamed while its first upload is on the wire goes back to its old name |
| B135 | A file just moved is sometimes not found by its new path |
| B131 | A write to a file whose create could not be read back lands as a conflict copy |
| B124 | After a replacing rename fails, the file it was replacing is gone from the mount |
| B123 | A second write open on a file just written fails with EIO |
| B122 | A rename across folders fails with EIO when the source folder holds the new name |
| B121 | A move right after a rename fails with EIO, and a rename across folders takes seconds |
| B120 | A file queued in a folder that has just landed is put off by five seconds |
| B119 | Reading a file just written through the mount needs the network |
| B118 | Creating a file or folder through the mount can fail with ENOENT although it was made |
| B112 | Folders shared by the same person disappear from Shared with me |
| B111 | Renaming a file back right after renaming it fails with EIO |
| B110 | A read that settles a provisional size deadlocks on its own page lock |
| B109 | A re-listing shrinks a file that is open for writing, and SQLite reads it as malformed |
| B108 | A file deleted during a mirror pass puts the sync folder in error |
| B107 | One read on a stalled connection hangs the reading program for half an hour |
| B106 | Saving over a file that was just read fails with "already exists" |
| B105 | A file deleted while its upload is in flight comes back as a conflict copy |
| B104 | Reading a file that was replaced by a rename fails with EIO |
| B103 | One stuck trash refresh freezes the trash listing until the daemon restarts |
| B102 | Removing a folder leaves the queued uploads of the files inside it |
| B101 | A shrunk file written past its end gets old bytes back, or fails with EIO; a write during a create's upload is lost |
| B100 | A freshly copied file reads as EIO or 53 bytes too big |
| B99 | Conflict copies in synced folders are logged again on every start and shown as "Resolved" |
| B98 | A superseded upload spins its drain thread forever and shows as "Uploading" at 0 bytes |
| B97 | Photo refreshes stack up on the control socket, and shutdown turns the resolve pass into a burst of "undecryptable photo" warnings |
| B96 | Staged bytes that name no node are never retired, and are re-announced on every boot |
| B95 | A real index.tmp is treated as a browser tempfile and parked forever |
| B94 | A parked create has no exit but the rename, so an abandoned one is stuck forever |
| B93 | A photo whose node would not resolve lost its name, and with it its tab and its group |
| B92 | Refreshing the gallery reported "Resource temporarily unavailable" |
| B90 | The local-file indexer walks into foreign FUSE mounts, and a dead one makes the daemon unkillable |
| B86 | An on-demand mount serves a listing that predates the mirror engine's own uploads |
| B85 | The read path assumes 4 MiB blocks; some revisions do not have them |
| B83 | An access-deferred op retries forever and is reported as an ordinary queued upload |
| B82 | A parent_uid cycle hangs the daemon on any short search |
| B81 | Prompt labels every Drive hit "My files", including device folders |
| B80 | 29% of the account is absent from the search index |
| B62 | Debian artifact omits required service and autostart units (HIGH-19) |
| B61 | Rotated single-use credentials can be lost (HIGH-18) |
| B60 | Config writes are non-atomic and parse errors are overwritten (HIGH-17) |
| B59 | Private state permission enforcement fails open (HIGH-16) |
| B58 | Control socket has unbounded frames and connections (HIGH-15) |
| B48 | Successful unlink can reappear from a stale remote listing |
| B42 | Local delete failure is recorded as sync success (HIGH-12) |
| B38 | Sync task panic could authorize destructive mode switch (HIGH-08) |
| B37 | Failed remote trash removed the local dentry (CRIT-04) |
| B34 | Writable POSIX exposure of read-only shared folders (MED-07) |
| B30 | SQLite write lock & mutex contention under heavy drain (MED-03) |
| B28 | Unhandled disk exhaustion (ENOSPC) in cache & scratch store (MED-01) |
| B27 | Orphaned write revision bricking offline sync queue (HIGH-07) |
| B22 | Synchronous block_on inside mutex guard (HIGH-02) |
| B19 | Remote folder trashing on mode switch failure (CRIT-02) |
| B14 | Provisional (ciphertext) sizes make rsync read short and abort the file |
| B13 | rename over an existing destination fails instead of replacing it |
Won't fix, not a bug, closed
| Entry | Title | What is left |
|---|---|---|
| B89 | The recorded SDK dependency is three minor versions out of date in the working memory | Closed: the file it described no longer exists |
| B87 | A revision's block geometry is learned a read too late | Won't fix: a bounded cost, accepted |
| B36 | Unicode normalization discrepancy (NFC vs NFD) (LOW-01) | Won't fix: NFC and NFD names stay distinct |
| B31 | Unimplemented special files & attribute modifications (MED-04) | Not a bug: by design |
| B3 | Not a bug: investigator error |
Fixed
| Entry | Title |
|---|---|
| B194 | On a fresh install the tray exits and the Takeout page cannot remember its archives |
| B192 | A new file closed empty reaches Drive only half a minute later |
| B181 | The app shows a queued edit kept as a conflict copy in English |
| B180 | A stop waits 10 s for a drain worker that slept through it |
| B169 | A stop lets a mirror pass go on until the runtime ends under it |
| B168 | A file removed with its folder while uploading is sent again, to the root |
| B167 | The mount takes seconds to come up on a large Drive |
| B166 | A folder drops or empties a file whose create lands while it is listed |
| B164 | A folder lists a new file twice while its create lands |
| B163 | A file read right after its create landed is empty |
| B162 | A folder removed and made again at once loses what goes into it |
| B161 | A file moved while it was being created shows up under its old name |
| B160 | A new folder reaches Drive only half a minute later |
| B133 | A file deleted while its upload is in flight can still come back as a recovered copy |
| B132 | A folder whose mkdir is still queued is gone after a restart |
| B130 | A read of a file that a remote change invalidates can hang the daemon |
| B129 | A file deleted or replaced while its create is on the wire stays on Drive |
| B128 | A file created offline over one deleted offline can be lost |
| B127 | A file created offline lands twice when the answer to its upload is lost |
| B126 | A file created offline under a name an offline rename freed lands as a conflict copy |
| B125 | A folder made offline fails with EIO once its listing is invalidated |
| B117 | An upload queued for a file pdfs move moved is listed by its raw id and put off once |
| B116 | Trashing a folder keeps the queued creates in its subfolders |
| B115 | pdfs ls shows a file's encrypted size |
| B114 | A name pdfs rm or pdfs rename took away still resolves for up to 30 seconds |
| B113 | A new file moved before its upload drains lands as a conflict copy of itself |
| B84 | A block read that comes back short truncates the file, silently |
| B79 | Every write in an on-demand device-folder mount fails EACCES |
| B78 | Profile backup fails: Cannot create file at the root of a device |
| B75 | Network work on fuser's single dispatch thread freezes the whole mount |
| B74 | A drained create never reaches a sync-folder mount's inode space, so the file reads as empty |
| B73 | Unimplemented FUSE handlers log a WARN per probe |
| B70 | Browser in-flight temp files uploaded and conflict-forked on the on-demand mount |
| B69 | Spurious (sync-conflict) copies from mtime-keyed revision identity (B16/B25 root fix) |
| B55 | Incomplete local scans are interpreted as deletions (CRIT-11) |
| B54 | Total-wipe guard does not protect a one-entry baseline (CRIT-10) |
| B53 | Superseding a pending operation is not transactional (CRIT-09) |
| B52 | Staged-write publication is not crash-safe (CRIT-08) |
| B51 | Scratch sidecar does not establish a durable recovery record (CRIT-07) |
| B50 | Failed orphan staging deletes the sole scratch copy (CRIT-06) |
| B49 | Failed conflict preservation still overwrites the local file (CRIT-05) |
| B47 | FUSE accepts path components longer than NAME_MAX |
| B46 | Combined cross-directory move and rename can fail out of date |
| B45 | Truncate after a queued rewrite corrupts or returns EIO |
| B44 | FUSE background workers have no coordinated shutdown (MED-10) |
| B43 | FUSE directory cookies are not stable (MED-09) |
| B41 | Sync folder removal races active reconciliation (HIGH-11) |
| B40 | Sync upload can stream torn live content (HIGH-10) |
| B39 | Sync download can overwrite a concurrent local edit (HIGH-09) |
| B35 | IPC Unix socket creation permission race (MED-08) |
| B33 | Lack of partial transfer resumption & head-of-line queue blocking (MED-06) |
| B32 | Race condition in concurrent handle release and open (MED-05) |
| B29 | Missing FUSE forget method / memory leak (MED-02) |
| B26 | Broken open-unlinked file semantics (HIGH-06) |
| B25 | Weak conflict baseline detection (mtime, size) (HIGH-05) |
| B24 | Directory hierarchy loop in rename (HIGH-04) |
| B23 | Un-fsynced scratch wipe on daemon restart (HIGH-03) |
| B21 | Untracked hole punching in fallocate (HIGH-01) |
| B20 | Un-fsynced temp file atomic rename in content cache (CRIT-03) |
| B18 | Subtree erasure via POSIX unlink / rmdir (CRIT-01) |
| B17 | FUSE mount lock contention, missing next_attempt_at in DB enqueue, and ioctl ENOTTY fix |
| B16 | aria2c preallocation / rapid sequential writes trigger false self-conflict copies |
| B11 | A file moved while its create is still queued reads as empty |
| B9 | Enter did not open the selected result in the launcher |
| B7 | Renamed directory reads as missing until the entry TTL expires |
| B6 | Daemon sets no file modes: control.sock is an unguarded authority |
| B5 | ls -l costs a network round trip per file (thumbnail xattr probes) |
| B4 | invalidate_listing silently skipped non-resident folders |
| B1 | FUSE rename loses the file (data loss) |
Acceptance harness
| Entry | Title |
|---|---|
| T1 | Managed acceptance raced daemon startup |
| T2 | Acceptance cleanup raced queued namespace replay |
| T3 | Mode-matrix preservation check raced mirror restoration |
| T4 | Successful managed cleanup left its own sentinel behind |
Entries
Newest first.
B210 — After a password reset, the mount retries "missing key" forever
Status: Fixed, not verified live
Found: 2026-10-15, issue #40: on a fresh 3.4.0 install the sign-in worked, but the daemon
logged mount failed: fetch My Files root: cryptography error: failed to decrypt: missing key
every 5 seconds. The app stayed on "Connecting…" because the daemon never got far enough to
create control.sock.
Where: ProtonDriveClient::ensure_my_files in proton-drive-rs; fetch_or_recall_root in
crates/pdfs-fuse/src/mount.rs
Repro: Reset the account password through Proton's recovery flow, which replaces the account keys, then sign in on a client without a cache.
Cause. A password reset gives the account new keys and locks the Drive volume whose share was
encrypted to the old ones. v2/shares/my-files still returns that share, so none of the unlocked
address keys can open its passphrase. The SDK only created a new volume when the share did not
exist, so every mount attempt failed the same way. The web client checks for an active regular
volume and creates a new one when there is none.
Fix. proton-drive-rs 0.9.1 handles a share key that fails to decrypt the way the web client
does. When GET volumes shows no active regular volume, it creates a new My Files volume and uses
that. The locked volume is left alone and can still be restored from the web with the previous
password.
Test: client::tests::a_locked_main_volume_is_not_an_active_regular_one,
client::tests::an_active_regular_volume_is_found_with_or_without_a_type (proton-drive-rs).
B209 — A blocked sign-in says the app version is no longer supported
Status: Fixed, not verified live
Found: 2026-10-13, issue #39: the first sign-in on a fresh 3.4.0 install showed "This version
of the app is no longer supported. Please update to continue using the app." Issue #28 showed the
same text on 3.0.0 to 3.1.2, and its debug log gave path="/auth/v4" http_status=422 code=2028.
Sign-in worked there after a day without attempts, with no change to the client.
Where: login_error_message in crates/pdfs-gui/src/app/pages/login.rs; cmd_login and
cmd_diagnose in crates/pdfs-cli/src/main.rs
Repro: Needs Proton to block the account or network, which cannot be triggered on purpose.
Cause. Proton answers a sign-in it blocks for abuse or rate limits with code 2028. Its Error
text claims the version is no longer supported. The SDK had no name for 2028, so it read as
Unknown, and the app passed Proton's text through. Users were told to update a current release.
pdfs diagnose also failed the database check on a fresh install, before the daemon had ever
started, which looked like part of the fault.
Fix. proton-sdk 0.8.3 names 2028 ResponseCode::Blocked. error::sign_in_blocked detects it,
and the app and pdfs login say that Proton is refusing sign-in for now and that waiting helps.
pdfs diagnose warns instead of failing when cache.db does not exist yet.
Test: error::tests::only_code_2028_is_a_blocked_sign_in,
pages::login::tests::a_blocked_sign_in_does_not_ask_for_an_update.
B208 — A file landing while its folder is pinned is kept but not available offline
Status: Fixed, not verified live
Found: 2026-10-11, CI simulation run of 3e4886f.
sim::daemon::tests::a_pinned_folder_keeps_every_file_under_it_until_it_is_unpinned waited
30 s for Docs/sub/b.txt to read as cached and gave up. It failed about once in 25 runs under load
and never alone.
Where: Core::pin_subtree in crates/pdfs-fuse/src/lib.rs
Repro: Pin a folder while a create in it lands, after the drain has repointed the tree at
the real uid and before it has read the node back. The file's cached blob must be gone, as after
pdfs cache clear, so the pin downloads it.
Cause. The pin walks the folder first and downloads each file under the modification time and size it read then. The landing's read-back then sets the time Drive gave the file, and keeps the uploaded bytes under it. The pin's download ended after that and stored the blob again under the old time. The cache records one time per blob, so the file read as stale: shown as kept offline, but downloaded again on the next read, and unreadable offline.
Fix. After each download the pin compares the file's time and size with the tree's. If they moved, it fetches the file again under the new ones, at most three times.
Test: sim::daemon::tests::a_folder_pinned_while_a_create_in_it_lands_keeps_that_file holds
the read-back and the pin's download with FakeClient::hold_next_download, new for this, and
lets them go in that order.
B207 — A move between locations goes ahead under a file opened just after its create landed
Status: Fixed, not verified live
Found: 2026-10-11, account acceptance run of 3.3.0, case "move refuses a file still being
written" on the on-demand → mirror pairing. pdfs move moved the folder while open.txt in it
was open for writing, and the write landed after the move.
Where: Core::has_pending_work_under in crates/pdfs-fuse/src/relocate.rs; retire_create
in crates/pdfs-fuse/src/drain.rs
Repro: Create a file in a mount, and open it for writing right after its create lands, before
the drain has repointed the tree at the real uid. Then pdfs move its folder to another
location. The account run did this by chance: it opened the file about a second after the
create landed.
Cause. retire_create takes the create out of the queue and re-keys its nodes row to the
real uid first, then repoints the in-memory tree and the open write handles. A file opened in
between gets a write handle that holds the placeholder uid. has_pending_work_under placed each
open handle by looking its uid up in the nodes table, which no longer has the placeholder, so it
found nothing open under the folder. The queue was already empty, and the move went ahead.
Fix. Open handles are placed in the in-memory tree by inode, which does not change when a create lands.
Test: sim::daemon::tests::a_move_out_of_an_on_demand_folder_refuses_a_file_still_open_for_writing
holds the drain at that point with FakeClient::hold_next_landing_of_any. The simulation can now
sync local folders: the fake Drive lists and registers devices, and the simulated daemons no longer
read the developer's own config.json.
B206 — An upgrade leaves the old daemon running, and nothing says so
Status: Fixed, not verified live
Found: 2026-10-11, issue #33. On
3.2.0 the reporter still saw Couldn't sync Documents: Read …/snapshots/37: Permission denied,
the B196 failure 3.2.0 fixed. Since 3.2.0 the scan leaves an unreadable subfolder out and
walk_local_dir can only fail the pass on the sync root, so that message comes from a daemon
older than 3.2.0.
Where: Response::Status and daemon_outdated in crates/pdfs-core/src/control.rs;
status_response in crates/pdfs-fuse/src/control.rs; the tray in crates/pdfs-gui/src/main.rs;
the sync card in crates/pdfs-gui/src/app/pages/status.rs; pdfs status and pdfs diagnose
Repro: With the service running, install a newer package, then open the app.
Cause. A package upgrade replaces the binaries but does not restart the user service, so the running daemon keeps the old release and its bugs until the next login or a manual restart. The app, the tray and the CLI start from the new binaries and could not tell, so they described the new release while the old one did the work.
Fix. The daemon reports its release in Status. A daemon from before the field omits it,
which reads as older. When the daemon is older than the front-end, the tray says "Update not
finished — restart Proton Drive" with a Restart Proton Drive item, the sync card says "Restart to
finish updating" with a Restart button, and pdfs status and pdfs diagnose print the
systemctl --user restart proton-drive command. A newer daemon is not flagged: then the
front-end is the stale one, and restarting the service would not help.
Test: control::tests::only_an_older_daemon_is_outdated;
tests::a_daemon_older_than_the_tray_needs_a_restart in pdfs-tray.
B205 — A photo upload the SDK refuses says "You're offline"
Status: Fixed, not verified live
Found: 2026-10-11, issue #38: uploading an image from the Photos page showed "You're offline"
while other uploads went through. Each upload_photo failed about 5 µs after it started, with no
request sent, and the journal gave no reason.
Where: CoreError::from_api and From<Error> for CoreError in crates/pdfs-core/src/error.rs;
the UploadPhoto handler in crates/pdfs-fuse/src/control.rs; the Photos page in
crates/pdfs-gui/src/app/pages/photos.rs
Repro: On an account without a photo library, open Photos and upload an image.
Cause. Every ProtonError other than Api was classified as ErrorKind::Offline, on the
reasoning that it never reached Proton. That holds for Transport only. InvalidOperation,
Crypto, KeysLocked and Serialization happen on this machine or after Proton answered. A
failure that fast is most likely the SDK's "account has no photos volume", which it returns from
its cache without sending a request. The app turns Offline into "You're offline", and the
daemon only sent the error back, so the journal did not say what it was.
Fix. proton_kind classifies by variant: Transport is Offline, InvalidOperation is
Invalid, Serialization is Remote, and Crypto and KeysLocked are Internal. The app then
shows "Couldn't upload photo" with the daemon's reason. The daemon logs a failed photo upload. The
Photos page greys out Upload when the daemon reports no photo library.
Test: error::tests::a_call_the_sdk_refuses_to_send_is_not_offline.
B204 — A write over a partial edit whose base was never read fails with EIO
Status: Fixed, not verified live
Found: 2026-10-09, by CI on 2a86830 (one_client_on_a_good_link, seed 1, which did not replay
locally). A partial write to x/f.txt, a file renamed twice before, was renamed over z/d. A
create over z/d a moment later failed with EIO, and the daemon logged "refusing write over
incomplete queued revision".
Where: ProtonFs::serve_open in crates/pdfs-fuse/src/filesystem.rs
Repro: Write at an offset into a file this client never read, then open it for writing again before the drain uploads the first write.
Cause. The release of a partial write fills its gaps from the content cache only. When the cache does not hold the base, the queued revision stays incomplete until the drain fills it from Drive, and a write open over an incomplete revision was refused. B123, B157, B169 and B170 each closed one way the cache could miss the base. Any other miss, such as a file never read here, still failed the next write open for the length of the drain debounce.
Fix. A write open over an incomplete queued revision copies the blob and fills its gaps from
the remote base with Core::fill_gaps, as the drain does before an upload. The open fails only
when that fill fails, for example offline.
Test: a_partial_write_over_bytes_never_read_takes_another_write_at_once in sim/daemon.rs.
Without the fix the second write fails with EIO.
B203 — pdfs unlock leaves a running daemon a spent refresh token
Status: Fixed, not verified live
Found: 2026-10-09, reading the code for issue #35 after the B201 fix
Where: crates/pdfs-core/src/auth.rs, stored_session_is_newer; crates/pdfs-fuse/src/link.rs,
Core::mark_session_expired; crates/pdfs-fuse/src/mount.rs; crates/pdfs-cli/src/main.rs,
cmd_daemon
Repro: With the daemon running, run pdfs unlock while the stored access token has expired.
Unlock resumes the same session, refreshes it and stores the new tokens. The daemon still holds
the old refresh token. If it refreshes before unlock restarts it, Proton refuses the spent token,
and since B202 the mount reports the session expired, although the keyring holds a live one.
Cause. Two processes ran the same session, and only one of them could hold its live refresh token. B201 stopped the daemon from writing its spent token over the live one, but nothing made the daemon pick up the live one.
Fix. Before a refused session counts as expired, auth::stored_session_is_newer compares the
keyring with what this process last stored. When the keyring holds a newer rotation of the same
session, or another login, the mount ends with MountOutcome::ResumeStored, and the daemon mounts
again at once from the keyring. At startup the same check turns the refusal into a mount error,
which the daemon retries. A rotation that another process already wrote into the blob is now
taken as this session's, so the daemon keeps writing its later rotations.
Test: auth::tests::a_session_rotated_by_another_process_is_newer_in_the_keyring and
auth::tests::a_rotation_already_in_the_blob_is_taken.
B202 — A refused session shows as offline forever, and nothing says to sign in again
Status: Fixed, not verified live
Found: 2026-10-09, issue #35: the tray and pdfs diagnose said "offline, 3 queued", and
diagnose also said "keyring session: ok", while every call in the log failed with
InvalidRefreshToken (http 400). pdfs sync now and restarting the service changed nothing.
Where: crates/pdfs-fuse/src/link.rs, Core::lost_link and Core::probe_until_online;
crates/pdfs-fuse/src/mount.rs, fetch_or_recall_root; crates/pdfs-fuse/src/background.rs;
crates/pdfs-core/src/control.rs, Response::Status; the tray, the app's status, pdfs status
and pdfs diagnose
Repro: Make the stored refresh token invalid (B201 is one way), then start the daemon. It mounts from the cache and says offline; the probe retries every 30 s, then every 5 minutes, and logs the refusal only at debug. A daemon that started online and loses its session later does not even go offline: its failures show up as EIO and drain backoff.
Cause. Every failure to fetch the root at startup counted as offline, and nothing told a refused session apart from a lost connection. No status field could carry it, so no front end could say it.
Fix. pdfs_core::error::session_revoked names the errors only a new login fixes:
InvalidRefreshToken, AccountDeleted, AccountDisabled. When the root fetch, the probe, the
drain, an event poll, the event seed or the photos root sees one, the mount goes offline with
session_expired set, logs one warning naming pdfs login, and stops probing and retrying.
Cached files stay readable and changes stay queued for the daemon the next login starts.
Response::Status carries session_expired. The tray shows "Session expired — sign in again"
with Sign In Again…, which opens the app's sign-in page; the app's status and Sync card say the
same, with a Sign In Again button. pdfs status and pdfs diagnose report it, the latter as
a FAIL. A refused session is not recorded as a sync issue of the op that ran into it.
Test: error::tests::only_a_refused_refresh_or_a_closed_account_revokes_the_session, the
tray's tests::a_refused_session_outranks_everything_else and
tests::a_daemon_that_does_not_say_has_a_live_session.
B201 — After signing in again, the old daemon puts its spent tokens back
Status: Fixed, not verified live
Found: 2026-10-09, reading the code for issue #35, where every call failed with
InvalidRefreshToken and signing in again did not obviously help.
Where: crates/pdfs-core/src/auth.rs, persist, register_refresh_handler, unlock and
migrate_legacy; crates/pdfs-cli/src/main.rs, cmd_daemon
Repro: With the daemon running, run pdfs login. It stores the new session and restarts the
unit. The old daemon's stop runs auth::persist, which wrote its own access and refresh tokens
into the blob, now the new session's. The new daemon starts with a refresh token of another
session, and the first refresh fails with InvalidRefreshToken. pdfs unlock had the same
shape: its session could rotate the tokens, and the restarted daemon's stop put the spent refresh
token back. unlock and migrate_legacy could do it themselves, writing a copy of the blob read
before verifying the mailbox password, which may have rotated the tokens.
Cause. Proton refresh tokens are single-use. store_tokens loaded whatever blob was stored
and replaced its tokens, without checking that the blob was still this session's.
Fix. Each session writes its rotations through a TokenSink. It writes only into a blob with
its session id that still holds the refresh token the session last left there; a blob that moved
on, to a new login or to another session's rotation, is left alone. The stop no longer writes the
session's tokens, only a rotation the keyring refused when it happened. unlock and
migrate_legacy read the blob again after verifying. A failed final write is an error in the log,
not a debug line claiming there was nothing to save.
Test: auth::tests::tokens_of_another_session_are_not_written,
auth::tests::tokens_are_not_written_over_a_newer_rotation_of_the_same_session,
auth::tests::a_sink_keeps_writing_its_own_rotations_and_stops_at_a_new_login.
B200 — Offline with changes queued, the daemon keeps several cores busy
Status: Fixed, not verified live
Found: 2026-10-09, issue #35: "offline, 3 queued", and systemd counted 17 minutes of CPU in
2 m 16 s of wall time, 10 hours over a longer run.
Where: crates/pdfs-fuse/src/drain.rs, Core::run_pending_drain and idle_wait;
crates/pdfs-core/src/db/ops.rs, Db::earliest_due_at
Repro: Queue a write, then cut the link (nmcli networking off). pidstat -t -p $(pgrep -x pdfs) 5 shows the drain threads using most of several cores for as long as the outage lasts.
Cause. Offline, a drain worker claims nothing, and then sleeps until the earliest queued op
falls due. Every queued op was due already, and with no attempts nothing moves its
next_attempt_at, so the sleep was zero. All 16 workers ran a query on every pass, the primary
one its idle chores as well. The same spin happened online when a due op sat behind a claimed op
on the same node, such as a rename queued behind a long upload: earliest_due_at counted it,
claim_next_due_op skipped it.
Fix. A worker that claims nothing, offline or paused, sleeps the full idle poll; a reconnect
or a resume wakes it. earliest_due_at now uses the claim's own exclusions, one SQL fragment for
both.
Test: drain::tests::an_offline_worker_waits_the_idle_poll_for_a_past_due_op,
db::tests::earliest_due_at_skips_an_op_held_behind_a_claimed_one.
B199 — A stopped share stays on the Shared by me page for minutes
Status: Partly fixed. The share itself is still not deleted.
Found: 2026-10-09, reported from the app: after Stop Sharing, the row stayed, and an
older share showed with no people and no link.
Where: crates/pdfs-fuse/src/sharing.rs, Core::stop_sharing_for_uid and
Core::list_shared_by_me; crates/pdfs-gui/src/app/pages/shared_by_me.rs, prompt_stop_sharing
Repro: Share a folder with a person, then choose Stop Sharing… on the Shared by me page. The row stays, now with "—" for people and link, until a reload some minutes later.
Cause. The Rust SDK has no call that deletes a share. Stop Sharing removes the public link, the
invitations and the members one by one, and leaves an empty share behind. Drive deletes an empty
share on its own, but only after a while, and until then GET v2/volumes/{vid}/shares still lists
it. The JS SDK ends unshareNode with DELETE drive/shares/{shareId}?Force=1 for exactly this
reason.
Fix. The daemon leaves a share with no member, no invitation and no link out of the Shared by
me listing, unless one of those lookups failed. The app also drops a stopped item from the page as
soon as the request succeeds, before the confirming reload. Left: a delete_share call in
proton-drive-rs, made at the end of stop_sharing_for_uid, so the share is really gone.
Test: none automated; the listing needs a live account.
B198 — Switching to online-only deletes local files that never reached Drive
Status: Fixed (unverified).
Found: 2026-10-08, reading the switch while fixing B196.
Where: crates/pdfs-fuse/src/devices.rs, Core::apply_sync_folder_mode;
crates/pdfs-fuse/src/sync.rs, mirror_tree_unsynced
Repro: Mirror a project folder with a .git/ in it, then switch it to Online only.
.git/ is gone from the disk and was never on Drive.
Cause. The switch evicts the whole local tree once the folder is idle. idle only says the
last pass applied everything it tried, and a pass never tries an ignored path (.git/ is ignored
by default), a symlink, or a folder it cannot read (B196). It also defers a file that is open for
writing without leaving idle, and a file can change after the pass ends. The eviction deleted
all of those.
Fix. Before the mode is written and anything is evicted, the switch walks the folder against the baseline, as a move out of a mirror folder already does, and refuses when something is only on this disk. The message names the path; an ignored one is called excluded from sync. A tracked file deleted here but not yet on Drive also refuses. An ignored one does not, since its baseline row stays for good and would hold the switch forever.
Test: unit tests a_fully_synced_mirror_folder_may_go_online_only,
an_ignored_folder_only_this_disk_has_holds_the_switch_back,
a_deleted_ignored_path_does_not_hold_the_switch_back,
an_edit_after_the_last_pass_holds_the_switch_back, a_new_file_at_the_root_holds_the_switch_back,
an_unreadable_folder_holds_the_switch_back and a_symlink_holds_the_switch_back. The switch
itself has not been driven against a live account.
B197 — Switching to online-only trashes synced files that became ignored
Status: Fixed (unverified).
Found: 2026-10-08, reading the push pass while fixing B196.
Where: crates/pdfs-fuse/src/sync.rs, Core::push_pass; crates/pdfs-fuse/src/sync/planner.rs,
local_deletions
Repro: Mirror a folder until build/ is on Drive, add build/ to .pdfsignore, then switch
the folder to Online only. build/ goes to the Drive trash, and the switch then deletes the
local copy.
Cause. The push pass that runs before the switch read its deletions as "in the baseline, not in the local walk". The walk skips ignored paths, so every synced path that later became ignored read as deleted. A full pass does not do this: it drops ignored paths from classification, which is what the user guide promises ("Ignoring never deletes anything").
Fix. The push pass reads its deletions from the same filtered baseline the wipe guard checks: without ignored paths, and without paths the scan could not read (B196).
Test: unit tests a_push_pass_never_reads_an_ignored_path_as_deleted and
local_deletions_come_out_shallowest_first.
B196 — One unreadable folder stops the whole synced folder from syncing
Status: Fixed (unverified). The reporter has not run the fix yet.
Found: 2026-10-08, #33. A synced
~/Documents held a restic cache a Docker container had created, which the user may not read.
Every pass failed with Couldn't sync Documents: read …/restic/cache/…/snapshots: Permission denied, and nothing else in Documents synced.
Where: crates/pdfs-fuse/src/sync.rs, walk_local_tree; crates/pdfs-fuse/src/sync/planner.rs,
Unscanned
Repro: In a mirror folder, sudo mkdir locked && sudo chmod 700 locked, then sync.
Cause. B55 made every scan failure fatal to the pass, because a path left out of the scan read
as a local deletion and was trashed on Drive. That was safe, but a folder the user may not read is
not transient: the pass failed on every poll, for good. The same held for a file deleted between
the listing and its stat, which a busy folder hits now and then, and for a name that is not UTF-8.
Fix. The scan leaves such a path out and records it, and the pass drops everything under it
from classification, the wipe guard and the push pass's deletions, as it already did for ignored
paths. A left-out path is then neither deleted on Drive nor downloaded into. A folder that lists
but does not allow stat is left out whole. An unreadable or absent root, and every other I/O
error, still fails the pass. The activity feed names the unreadable paths once each time the set
changes, and the folder stays idle. Adding the folder to .pdfsignore was already a workaround
and still silences the report.
Test: unit tests an_unreadable_subdirectory_is_left_out_and_the_rest_scans,
a_listable_but_unsearchable_folder_is_left_out_whole,
a_path_the_scan_left_out_is_never_a_local_deletion,
unscanned_covers_a_path_and_what_is_under_it_but_not_its_siblings and
a_non_utf8_name_is_left_out_and_reported. an_unreadable_root_fails_the_scan and
an_absent_root_fails_the_scan still hold.
B195 — The app crashes when Proton asks for a CAPTCHA on secureblue
Status: Fixed (unverified). On 2026-10-08 the reporter confirmed that the app survives, but
the CAPTCHA still could not be solved; the browser fallback below has not run on secureblue yet.
Found: 2026-10-07, #29. On
secureblue 44 (a hardened Fedora Atomic), pdfs-app aborted a few seconds after the sign-in, with
bwrap: Can't get type of source /run/systemd/journal/socket: Permission denied and
Failed to fully launch dbus-proxy.
Where: crates/pdfs-gui/src/app/pages/verify.rs, crates/pdfs-gui/src/app/pages/login.rs,
crates/pdfs-core/src/auth.rs, crates/pdfs-cli/src/main.rs
Repro: Sign in on secureblue from a network Proton asks a CAPTCHA for, such as a VPN.
Cause. The CAPTCHA page runs in a WebKitGTK view (B8). Before it loads anything, WebKit starts
its sandbox: bubblewrap, with the journal socket bound in, and xdg-dbus-proxy. secureblue keeps
bubblewrap from the socket, the proxy does not start, and WebKit ends its process with g_error.
WebKit binds the socket with --ro-bind-try, but bubblewrap skips only a missing path, not one it
may not read, and WebKit has no switch to leave the socket out.
The view was in the app's own process, so the whole app went down, and nothing in our code can
catch that.
Fix. The view runs in a child, pdfs-app --human-verification, that reads the URL on stdin
and prints the token on stdout. A child that exits without a token after a crash or a non-zero
status no longer ends the sign-in: the app offers the page in the user's own browser, as Proton
Mail Bridge does. The tab cannot post the token back, so the user says when they are done, and the
retry carries the challenge token, typed with every offered method. pdfs login takes the same
path. The sandbox is not switched off by the app: users choose a hardened system for that
isolation. TROUBLESHOOTING.md keeps the one-time workaround as a last resort.
Test: unit tests a_window_that_died_fails, a_window_closed_without_a_token_cancels,
a_token_counts_even_if_the_window_died_after_it, the_token_is_found_among_other_output,
the_browser_gets_the_full_verification_page and
a_browser_verification_retries_with_the_challenge_token. The browser retry has not run against a
live challenge: Proton only asks for one from a network it distrusts.
On 2026-10-08 the child ran under real WebKit on Arch against a local page that posts a completion,
and handed back its token.
B194 — On a fresh install the tray exits and the Takeout page cannot remember its archives
Status: Fixed.
Found: 2026-10-07, in the log of #29:
failed to bind tray single-instance socket: No such file or directory and
cannot remember staged Takeout archives: No such file or directory.
Where: crates/pdfs-gui/src/main.rs, build_window in crates/pdfs-gui/src/app/main.rs
Repro: Start pdfs-app with no ~/.local/state/proton-drive-linux, as on a fresh install. The
tray is gone, and staging a Takeout archive logs the warning.
Cause. Both write to the state directory, and only the daemon and the CLI made it
(AppDirs::ensure). The app starts the tray before the first sign-in, so before the daemon has
ever run.
Fix. pdfs-tray and pdfs-app call AppDirs::ensure at start, which also makes the
directories owner-only (B6).
Verified: 2026-10-08, pdfs-tray with XDG_STATE_HOME, XDG_CONFIG_HOME and
XDG_CACHE_HOME pointing at an empty directory. It exited with the error before the fix, and
afterwards ran with tray.sock in a 0700 state directory.
B193 — Deleting a file whose create failed can trash a file of the user's on Drive
Status: Fixed (unverified).
Found: 2026-10-05, in the confirmation run on 1436c5f (one_client_on_a_flaky_link seed 6;
no replay reproduced it). A file renamed over y/y/d, then emptied, was gone from Drive and
from the local database, with nothing queued. The run kept no log. The path below explains it,
and the test reproduces that path every time.
Where: Core::withdrawn_twin_in in crates/pdfs-fuse/src/drain.rs.
Cause. A delete turns a create that was sent into a trash of what it may have made, under a
name it was sent with (B151, B184). The drain looks there for a node made in the last two
minutes that holds bytes the create sent, or none. It did not leave out a node in the tree, so
an empty file of the user's renamed to that name matched, and was trashed. A create's own search
for its twin (Core::adoptable_twin) already left such a node out, and nodes this mount trashed
(B162).
Fix. The trash's search leaves out the same nodes. A file a lost answer made that a listing has since put in the tree stays on Drive, where the user can see and delete it.
Test: simulation test a_withdrawn_create_leaves_the_file_renamed_to_the_name_it_was_sent_under
(failed before the fix: Drive was empty).
B192 — A new file closed empty reaches Drive only half a minute later
Status: Fixed. Found: 2026-10-05, while writing the test for B193. An empty file made on an idle mount was still queued 30 s later.
Where: Core::queue_revision in crates/pdfs-fuse/src/lib.rs.
Cause. Queueing a file's create does not wake the drain: its bytes ride on the create, and
the release that attaches them wakes it (B160). A file closed without a write (touch, : > f)
attaches nothing, so its create waited for the workers' 30 s idle poll.
Fix. Closing a clean handle of a file whose create is queued wakes the drain.
Test: simulation test a_file_made_empty_while_the_drain_is_idle_reaches_drive_at_once
(failed before the fix: the file was still queued after 10 s).
B191 — A file deleted after its create failed and was written again can stay on Drive
Status: Fixed (unverified).
Found: 2026-10-05, in the confirmation run on 1436c5f (one_client_on_a_flaky_link seed 11,
one replay in six). A create of a.txt got no answer, though Drive had made the file. The file
was written again, renamed twice and deleted. Drive kept a.txt, and nothing was queued.
Where: drop_doomed_ops in crates/pdfs-core/src/db/ops.rs.
Cause. A delete turns a create that was sent into a trash of what it may have made (B151).
Whether it was sent was read from its attempts and its claim. A write after the failure attaches
its bytes to the create and clears the attempts, so that the retry goes out at once
(Db::attach_blob_to_create). The delete then took the create for one never sent, and dropped
it whole.
Fix. A create with a target recorded (sent_to, B184) counts as sent as well. Test
a_create_rewritten_after_it_failed_still_leaves_a_trash_when_deleted in
crates/pdfs-core/src/db/tests.rs.
B190 — A file saved just before the daemon stops can keep its old content on Drive
Status: Fixed (unverified).
Found: 2026-10-05, in the confirmation run on 2baad67 (one_client_on_a_flaky_link seed 9,
about one run in three). A client wrote a 32-byte file and closed it, and the daemon was
restarted at once. Drive held the file empty, and nothing was queued.
Where: the stop in crates/pdfs-fuse/src/mount.rs, and Core::queue_unreleased_writes in
crates/pdfs-fuse/src/lib.rs.
Cause. close(2) returns before the kernel sends release, and the release is what queues
a handle's bytes. The stop aborts the FUSE connection so that the unmount cannot hang, and the
abort drops the requests the kernel still has queued. A release among them never reached the
daemon. The file's create went out with what it held before the write, and the bytes went with
the handle's scratch file. A handle a program still held open at the stop lost what it had
written in the same way.
Fix. Once every session has ended, the stop queues each write handle the kernel never
released, as its release would have. Test a_stop_keeps_what_a_handle_wrote_before_its_release
in crates/pdfs-fuse/src/sim/daemon.rs.
B189 — A change that failed on a short outage waits out its retry long after the link is back
Status: Fixed (unverified).
Found: 2026-10-05, in the confirmation run on 2baad67 (three_clients_one_writer_each
seed 6, every replay). A client's link was down for 0.3 s, during which it made a folder and a
file. Their uploads failed once, and 4 s after the link came back they were still queued.
Where: the failure path of Core::run_pending_drain in crates/pdfs-fuse/src/drain.rs.
Cause. Every syscall and the event poll take the mount offline when Drive does not answer, and the probe that brings it back makes every backed-off op due at once. The drain was the one caller that did not. An outage that only the drain met left the mount online, so nothing probed for the link's return, and the ops sat out a backoff that doubles with each failure, up to five minutes.
Fix. A drain attempt that cannot reach Drive takes the mount offline like any other call.
Test a_write_that_failed_on_a_short_outage_goes_out_soon_after_the_link_returns in
crates/pdfs-fuse/src/sim/daemon.rs.
B188 — A restart can leave a queued write retrying forever on a file deleted while open
Status: Fixed (unverified).
Found: 2026-10-05, in the confirmation run on 2baad67 (three_clients_one_writer_each
seed 12). With its link down, a client overwrote a file, deleted it while open, appended to it
and restarted. After the restart a revision of the deleted file retried on "No such file or
directory" until the run gave up.
Where: Core::reconcile_staging in crates/pdfs-fuse/src/lib.rs.
Cause. Since B186's fix, the queued trash of an open file keeps the staged bytes the handle reads, and the trash's landing removes them. A restart in between forgets which handle the bytes were kept for. The scan at mount found them owned by no op and queued them again as a revision of the node. The trash, queued first, landed and removed the bytes, and the revision kept retrying on a blob that was gone. B186's fix brought this in, so no release has it.
Fix. The scan at mount discards staged bytes whose node has a queued trash, as the trash would
have. Test a_restart_does_not_bring_back_the_write_of_a_file_deleted_while_open in
crates/pdfs-fuse/src/sim/daemon.rs.
B187 — An uploaded file can leave a copy of itself in the cache until the next start
Status: Fixed (unverified).
Found: 2026-10-05, writing the test for B186: after its file was deleted, the cache still
held a .tmp file with its bytes.
Where: ContentCache::store_file in crates/pdfs-core/src/cache.rs.
Cause. The cache adopts an uploaded file by linking its staged bytes to a temp name and renaming that over the cached blob. A landed create and the revision after it can adopt the same staged bytes. The temp name was then a second name of the cached blob, and a rename between two names of one file does nothing. The temp file stayed, outside the cache's budget, until the next start swept it, even after the file was evicted or deleted.
Fix. The temp name is removed after the rename. Test
store_file_twice_from_the_same_blob_leaves_no_temp_file in crates/pdfs-core/src/cache.rs.
B186 — A file deleted offline while open fails a partial write with "Input/output error"
Status: Fixed (unverified).
Found: 2026-10-05, in the confirmation run on 3073052 (three_clients_one_writer_each
seed 4; it fails on da013c2 too). With its link down, a client opened a file it had uploaded,
deleted it, and appended to it. The append took 30 s and failed with EIO.
Where: Core::queue_trash in crates/pdfs-fuse/src/lib.rs and Core::drain_trash in
crates/pdfs-fuse/src/drain.rs.
Cause. A delete sent to Drive at once keeps what an open handle reads, its queued and cached bytes, until the final close (B106). A delete queued, as local-first and offline do, dropped them at once. A partial write fills its gaps from those bytes, so it had to download the file, and offline it could not.
Fix. A queued delete of an open file keeps them as well. The final close lets go of them
once the delete has landed, and the delete landing lets go of them once the file is closed.
Sim test a_file_deleted_offline_while_open_still_reads_and_takes_writes.
B185 — A new file written again just as it starts uploading waits out a retry
Status: Fixed (unverified).
Found: 2026-10-05, in the confirmation run on 3073052 (one_client_on_a_good_link seed 6,
then one_client_deleting_files_as_they_upload seed 4). The queues did not drain in time: a
create had failed with "No such file or directory" and was backing off.
Where: Db::retry_if_retargeted in crates/pdfs-core/src/db/ops.rs, called from the drain
in crates/pdfs-fuse/src/drain.rs.
Cause. A write to a file whose create is queued attaches a new blob to the create and discards the old one at once. An attempt that had claimed the create but not yet opened the old blob failed on a file that was gone, and backed off as if Drive had refused it. The new blob was ready, but the create waited out the backoff.
Fix. A failed create whose blob was replaced while it ran is due again at once, the way one
renamed while it ran already was. Test renaming_a_backing_off_create_makes_it_due in
crates/pdfs-core/src/db/tests.rs.
B184 — A file renamed after its create lost its answer stays on Drive under its old name
Status: Fixed (unverified).
Found: 2026-10-05, replaying one_client_on_a_flaky_link seed 11 with the B183 fix. a.txt's
create reached Drive and lost its answer. The file was then rewritten, renamed to b.txt and
then d, and deleted. Drive kept a.txt with the bytes the create had sent.
Where: Core::withdrawn_twin and Core::create_local_node in
crates/pdfs-fuse/src/drain.rs.
Cause. A rename while a create waits rewrites the op's parent and name, but the node a lost answer left stays where the create was sent. The trash a delete made of the create, and the create's retry, looked for that node under the op's last name only. The trash found nothing and left the file on Drive. The retry made a second file under the new name, so Drive had both.
Fix. Each attempt records the parent and name it is sent under on the op's row, in a new
pending_op.sent_to column (schema 42). The trash looks for its node under each of them. A
retry first adopts a node an earlier attempt made under another name, and the rename to the
current name follows it. Sim tests a_file_renamed_after_its_create_lost_its_answer_lands_once
and a_file_renamed_and_deleted_after_its_create_lost_its_answer_leaves_nothing, and
what_a_create_sent_outlives_a_newer_blob_its_trash_and_a_restart in
crates/pdfs-core/src/db/tests.rs.
B183 — A file deleted after a restart stays on Drive when its create lost its answer
Status: Fixed (unverified).
Found: 2026-10-05, by the known-bug confirmation run (PDFS_SIM_KNOWN=fail) on da013c2:
one_client_on_a_flaky_link seed 11 failed every replay. d was made, written again while
its create was on the wire, and deleted after several restarts. Drive kept a d with the bytes
the create had sent, and the next file renamed to d landed as d (sync-conflict …).
Where: Core::withdrawn_twin and Core::adoptable_twin in crates/pdfs-fuse/src/drain.rs.
Cause. The trash a delete makes of a create that was sent looks for a node Drive made with the bytes the create sent (B151). Those were kept in memory only (B156). After a restart the trash knew only the blob the op held last, which a later write had replaced. It found no node of its own and left the file on Drive. A create retried after a restart had the same gap and forked a conflict copy.
Fix. Each attempt records the SHA-1 of the blob it sends on the op's row, in a new
pending_op.sent_sha1 column (schema 42). The row keeps it through a newer blob, through the
turn into a trash and across restarts. The in-memory map is gone. Tests
what_a_create_sent_outlives_a_newer_blob_its_trash_and_a_restart in
crates/pdfs-core/src/db/tests.rs, and sim test
a_file_deleted_after_a_restart_leaves_drive_with_what_its_create_sent.
B182 — A write after a restart over a create never read back lands as a conflict copy
Status: Fixed (unverified).
Found: 2026-10-05, by the known-bug confirmation run on da013c2:
one_client_on_a_flaky_link seed 6 failed two of three replays. y/x/c.bin's create landed and
the link went down before it was read back. The daemon restarted, and a write to the file landed
as c (sync-conflict …).bin, with Drive keeping the old bytes.
Where: Core::finish_adoption in crates/pdfs-fuse/src/drain.rs.
Cause. A create that landed but was not read back is adopted before the next op for the node is sent (B143). The node is remembered for that in memory, and a restart forgot it. The tree then still had the file as it was made here, with no revision of Drive's, and a write was based on that. The drain found Drive's revision in its way and took it for someone else's.
Fix. A queued write over a file the tree holds without a revision id is adopted first, as
B143 does, after a restart too. Sim test
a_write_after_a_restart_over_a_create_never_read_back_is_not_a_conflict.
B181 — The app shows a queued edit kept as a conflict copy in English
Status: Fixed.
Found: 2026-10-05, checking that every name collision and conflict the drain resolves is
reported (sync issues, docs/ARCHITECTURE.md §5). Each one is in the activity log, and the app words
each in the user's language but one.
Where: detail_label in crates/pdfs-gui/src/app/pages/activity_text.rs; the daemon
writes the entry in Core::keep_as_conflict_copy and Core::abandon_to_staging
(crates/pdfs-fuse/src/drain.rs).
Cause. A queued edit whose file was changed, trashed or removed on Drive meanwhile is uploaded as a conflict copy, or kept in staging when it cannot be. The entry's detail is the reason followed by "; local changes uploaded as {copy}" or "; local changes kept at {path}". The app knew neither shape, so it showed the detail as the daemon wrote it, in English: "The remote revision changed under the queued write (based on revision …); local changes uploaded as …".
Fix. The app reads both shapes and the three reasons the daemon gives, and says each as one sentence: "It was changed on Proton Drive, so your changes were uploaded as {name}". A reason it does not know still gets a sentence without the why. Translated in every catalog.
Test: an_edit_kept_as_a_copy_says_why_in_words in
crates/pdfs-gui/src/app/pages/activity_text.rs.
B180 — A stop waits 10 s for a drain worker that slept through it
Status: Fixed.
Found: 2026-10-04, by the simulation suite on 0f3c06f: a_stop_ends_a_mirror_pass_after_the_uploads_it_is_sending
found "background workers did not finish in time" in the log, left there by two earlier tests
whose stops had given up on a drain worker. On its own,
a_folder_made_again_while_the_old_one_is_trashed_is_a_new_folder took 11 s instead of 2 s in
about one run in three.
Where: Core::wait_for_drain_work and Core::wait_for_drain_work_or_due in
crates/pdfs-fuse/src/drain.rs.
Cause. The drain workers sleep on one flag and one condvar. A stop sets its flag, then raises the wake-up and wakes them all, and the first worker to wake takes the wake-up. A worker that was between its stop check and its sleep at that moment found no wake-up and slept its timeout, up to 30 s. The stop gave up on it after 10 s and shut down around it. With 16 workers since step 7 this was common.
Fix. A worker checks the stop under the wake-up's lock before it sleeps. The stop sets its flag before it takes that lock, so a worker that finds it unset is asleep when the wake-up comes.
Test: a_drain_worker_that_missed_the_stop_wake_up_does_not_sleep in
crates/pdfs-fuse/src/drain.rs.
B179 — A file read just as its create lands reads empty
Status: Fixed (unverified).
Found: 2026-10-04, by the account run on 378667f, on on-demand sync folders: the unusual
names case read <angle>|pipe back as 0 bytes instead of 12, and B70's completed download as 0
bytes instead of 147456. Both files had a write that landed during their create, which the
daemon queued as a revision.
Where: Core::read_range in crates/pdfs-fuse/src/reads.rs.
Cause. A read takes the file's uid from the tree and reads on a transfer worker. A file made on the mount has its stand-in uid there until its create lands. When it lands, the queued write moves to the uid Drive gave the file. A read that still asked by the stand-in found no write queued under it, took the file for one created and never written, and returned nothing. The read waits on the transfer lane, which a run keeps busy, so the window is not small.
Fix. A stand-in with no queued write is looked up by its local id. Once its create has landed, the read goes on under the uid Drive gave it.
Test: a_file_read_as_its_create_lands_has_its_bytes in crates/pdfs-fuse/src/sim/daemon.rs.
B178 — Trashing a folder while sync is paused keeps a create in it as a trash
Status: Fixed (unverified).
Found: 2026-10-04, by the account run on 378667f, on an on-demand sync folder: B116's
pdfs rm b116 with sync paused left the create of b116/sub/deeper/b116 held.bin queued as
trash #16098.
Where: Core::run_pending_drain in crates/pdfs-fuse/src/drain.rs.
Cause. Trashing a folder drops the creates queued below it, but keeps a create that may already be on the wire as a trash, for a node Drive may have made (B151). A claimed create counts as such. A drain worker that woke while sync was paused or offline claimed the next due op and only then saw it could not send it, and handed it back. The rename that lets a parked create go wakes the workers, so a trash in that moment kept a create nothing had sent.
Fix. A worker checks that it is online and not paused before it claims anything.
Test: a_paused_drain_claims_nothing in crates/pdfs-fuse/src/lib.rs.
B177 — A folder listed just as its create lands answers "Input/output error"
Status: Fixed (unverified).
Found: 2026-10-04, by the account run on 378667f, on an on-demand sync folder: B100's
lstat of b100-copy/dir-0/sub-0/file-00.bin failed with EIO. The daemon logged
enumerate folder children failed folder_uid=local~163903 ... InvalidEncryptedIdFormat 85 ms
before dir-0's create landed.
Where: Db::children_if_listed and Db::set_listed in crates/pdfs-core/src/db/nodes.rs.
Cause. A folder made on the mount is marked as listed, so its listing comes from the database and never from Drive, which has not heard of it. The flag was looked up by uid. In B173's gap the row already has the uid Drive gave the folder and the tree still asks by the stand-in, so the lookup found no row, and the listing went to Drive with the stand-in.
Fix. A stand-in finds its row by its local id, before and after its create lands.
Test: a_stand_in_folder_stays_listed_once_its_create_has_landed in
crates/pdfs-core/src/db/tests.rs.
B176 — A write closed just as its file's create lands is dropped
Status: Fixed (unverified).
Found: 2026-10-04, by one_client_on_a_good_link (seed 12) on 378667f, run three at a time
in a loop. A write at offset 6 to a file whose create was landing never reached Drive; Drive
ended with the file's first bytes and the later truncate only.
Where: Core::queue_revision in crates/pdfs-fuse/src/lib.rs.
Cause. A write handle closed on a file not yet on Drive first checks whether the create has landed and, if so, follows it to the uid Drive gave it. It then checks that the create is still queued; a file with neither was unlinked before it was created, and its write is dropped. A create that landed between the two checks looked like that, and the write was dropped.
Fix. The landing removes the create op and gives the row its uid in one transaction. So a missing create op is checked against the row, and a landed file's write is queued for its uid.
Test: a_write_closed_as_its_create_lands_is_queued in crates/pdfs-fuse/src/lib.rs.
B175 — A file replaced by a rename just as its create lands stays on Drive
Status: Fixed (unverified).
Found: 2026-10-04, by one_client_on_a_good_link (seed 8) on 378667f, run three at a time
in a loop. rename b.txt -> a.txt replaced an a.txt whose create had landed a few
milliseconds before, and the next rename a.txt -> d moved the old a.txt on Drive. Drive
ended with the old bytes in d and the new ones still in a.txt, and the queue did not drain.
Where: Core::remove_replaced in crates/pdfs-fuse/src/lib.rs.
Cause. The same gap as B173, through rename. The replace found the old file under its
stand-in, found no op left to drop, and took it for a file Drive had never seen. So nothing
trashed it on Drive, and the tree took it back under its name once it got its uid.
Fix. As in B173: after dropping a stand-in's ops, the replace checks whether the node's row has a real uid. If it has, the file is trashed on Drive like any other replaced file.
Test: a_file_replaced_as_its_create_lands_is_trashed_on_drive in
crates/pdfs-fuse/src/sim/daemon.rs.
B174 — A file renamed just as its create lands is refused with "Permission denied"
Status: Fixed (unverified).
Found: 2026-10-04, by the simulation suite on 378667f (one_client_on_a_good_link, seed 6):
rename c.bin -> b.txt answered EACCES right after a settle. Not reproduced in three replays.
Where: Db::effective_node_access in crates/pdfs-core/src/db/share_access.rs.
Cause. Every change on the mount first asks whether the node may be written, and a node with no row is refused. When a create lands, its row takes the uid Drive gave it before the tree does (B173's gap). A rename in between asked by the stand-in the tree still had, found no row under it, and was refused.
Fix. A stand-in is looked up by its local id, so it finds its row before and after the create lands.
Test: a_stand_in_keeps_its_access_once_its_create_has_landed in
crates/pdfs-core/src/db/tests.rs.
B173 — A file deleted just as its create lands stays on Drive
Status: Fixed (unverified).
Found: 2026-10-04, by one_client_on_a_good_link (seed 12) on d850265, run three at a time
in a loop: one of nine runs failed. A deleted c.bin was still on Drive with its old bytes,
and the next c.bin made on the mount landed as a conflict copy beside it.
Where: trash_child in crates/pdfs-fuse/src/filesystem.rs.
Cause. When a create lands, the drain retires its op and moves the node's row to the uid Drive gave it, then gives the tree that uid. In between, it caches the uploaded bytes. A delete in that gap still found the file under its stand-in, found no op left to drop, and took it for a file Drive had never seen. So nothing trashed it on Drive.
Fix. After dropping a stand-in's ops, the delete checks whether the node's row has a real uid. If it has, the create landed first, and the file is trashed on Drive like any other.
Test: a_file_deleted_as_its_create_lands_is_trashed_on_drive in
crates/pdfs-fuse/src/sim/daemon.rs, which holds the drain between the two steps.
B172 — A file renamed while its upload is read back shows its old name
Status: Fixed (unverified).
Found: 2026-10-04, by one_client_on_a_good_link (seed 8) on 2502af6, run three at a time
in a loop: two of nine runs failed. In one, a write to f.txt answered ENOENT; in the other,
a rename of f.txt. The mount showed the file as a.txt again.
Where: refresh_after_upload and place_landed in crates/pdfs-fuse/src/drain.rs.
Cause. a.txt was renamed to f.txt while its create was on the wire, with a write behind
it. The create landed, and its op became a revision with a rename queued after it. The upload
retires its op before it reads the node back, so the rename went out while Drive answered the
read-back, and that answer still had the old name. The drain asked whether a rename was queued
only after the answer came, found none, and put a.txt back into the tree. The create's
read-back already asks first (B161); the upload's did not.
Fix. The upload's read-back notes whether a rename is queued, and where the tree has the file, before it asks Drive. The tree keeps the name and folder the user gave the file if a rename was queued then or is now, or if the file moved meanwhile.
Test: a_file_renamed_while_its_upload_is_read_back_keeps_its_new_name in
crates/pdfs-fuse/src/sim/daemon.rs.
B171 — A file deleted while its upload fails stays on Drive for the upload's backoff
Status: Fixed (unverified).
Found: 2026-10-04, by the simulation suite (one_client_on_a_good_link, seed 9). The queue
did not drain within 4 s of the settle: it still held a trash of b.txt with one attempt and
the error "No such file or directory (os error 2)", due seconds later.
Where: drop_doomed_ops and Db::record_op_failure in crates/pdfs-core/src/db/ops.rs, and
the drain loop in crates/pdfs-fuse/src/drain.rs.
Cause. A file deleted after its create was sent keeps its op as a trash, in case the create made the node and only its answer was lost (B151). The op kept the create's attempts, error and backoff. The drain also recorded the failure of a create attempt still on the wire on the row, which by then was the trash. Here the attempt failed because a write had replaced the blob it was reading, and the trash waited out the create's backoff.
Fix. The trash starts with no attempts and is due at once, as a create that becomes a
revision already does. The drain records a failed attempt only if the op is still the kind it
ran as (Db::record_attempt_failure).
Test: a_create_deleted_after_it_was_sent_leaves_a_trash_that_is_due_at_once in
crates/pdfs-core/src/db/tests.rs.
B170 — A write opened as an upload lands fails the next write with EIO
Status: Fixed (unverified).
Found: 2026-10-04, by CI on 9bbc994 and by three parallel copies of
one_client_on_a_good_link (seed 4; seed 5 in the local suite). A write at an offset was
refused with EIO, and the daemon logged "refusing write over incomplete queued revision".
Just before, an open had retried with "queued revision changed while opening for write", and
the queued write it then made was incomplete.
Where: serve_open in crates/pdfs-fuse/src/filesystem.rs, and Core::refresh_after_upload
in crates/pdfs-fuse/src/drain.rs.
Cause. When an upload lands, the drain caches its bytes under the node as Drive has it, rebases the open write handles onto it, lets go of the blob, and then puts the node in the tree. A write open in between found no blob to copy and took the base it had sampled at its start: the mtime and size the tree had before the upload. The release of its partial write filled the gaps from the cache under that key, found nothing, and queued an incomplete revision, and a write open over one is refused (B123). Under load, the open retried because of the landing and met that window.
Fix. From before the drain lets go of the blob until the tree has the node, the landing node is kept aside. A write open with no queued revision takes its base from that node, or else from the tree as it is when the handle goes in, not from its first sample.
Test: a_write_opened_as_an_upload_lands_takes_the_landed_revision_as_its_base in
sim/daemon.rs holds the drain in that window with the fake Drive's new hold_next_landing,
writes at an offset, and writes again after it. Without the fix the second write fails with
EIO.
B169 — A stop lets a mirror pass go on until the runtime ends under it
Status: Fixed.
Found: 2026-10-04, in the journal of the account run on the working tree after B165. A
stop that came while a mirror pass was uploading logged "background workers did not finish in
time" for pdfs-sync, then "sync: task panicked … was cancelled" and a run of "no remote
parent" warnings: the pass had gone on to deeper folders while the stop waited for it, and the
runtime's end cancelled the uploads under it.
Where: flush_batch, walk_remote and reconcile_pass in crates/pdfs-fuse/src/sync.rs.
Cause. The mirror pass never asked whether the daemon was stopping. It sends its work a level of folders at a time and went on to the next level after the stop, until the stop gave up on it at its 10 s limit (B165).
Fix. Once the daemon is stopping, a batch starts none of its ops still waiting for their turn, lets the ones already sending finish, and ends the pass. So do the remote walk and the start of a pass. The folder keeps the state it had before the pass; the next start's pass picks up from the baseline.
Test: a_stop_ends_a_mirror_pass_after_the_uploads_it_is_sending in sim/daemon.rs holds a
mirror upload, stops the daemon, then lets the upload through, and checks that the file in the
deeper folder was not sent and the stop did not run into its limit.
B168 — A file removed with its folder while uploading is sent again, to the root
Status: Fixed. Found: 2026-10-04, by the journal check of the account run on the working tree after B165: "re-interned an authority missing from the local tree; retrying uid=local~151607". The wide directory case removed its folder while three of its files were still on their way to Drive. The folder's trash landed first, and Drive refused the three creates (422). The drain then logged "parent folder is gone remotely; creating in the root instead" for each, and failed only because the deleted files' bytes were gone. The trashes the deletes had queued then found the folder missing from the tree and read it back from Drive.
Where: create_local_node and resolve_unknown_authority in
crates/pdfs-fuse/src/drain.rs.
Cause. A delete of a file whose create is on the wire turns the create's op into a trash, which later looks in the create's folder for a node the create may have made (B151). The create's own failure path did not look at that: a create refused because its folder is gone is re-homed in the root (for a folder trashed on another device), deleted file or not. And the trash's access check asks about that folder, whose row the folder's own trash had removed, so it read the trashed folder back into the cache, deferred, and only then found nothing to trash.
Fix. A create a delete withdrew is not re-homed. A withdrawn create's trash whose folder is trashed or gone is done: Drive took the node it was looking for along with the folder. It used to record a sync issue when the folder was gone for good.
Test: a_file_removed_with_its_folder_while_uploading_stays_removed in sim/daemon.rs
holds the file's create on its way to Drive until the folder's trash has landed, then checks
that nothing was sent to the root and the folder was not read back (each fails without its
half of the fix). FakeClient::hold_next_create is new for it.
B167 — The mount takes seconds to come up on a large Drive
Status: Fixed.
Found: 2026-10-04, looking into the cleanup of the account run on the working tree after
B165, which found no control socket after the last daemon restart. Every start in that run
took 2 to 3 s from "mounting" to "hydrated metadata cache from db" (37,000 nodes), the one
before the cleanup 8.8 s. That was not the cleanup's failure: the socket is bound before the
hydrate, and a request made during it waits. The cleanup ran some 1.5 s after the restart,
before the new daemon got as far as binding it. pdfs status reports a daemon that is not
listening yet as not running, and the run's queue wait took that for an empty queue.
Where: hydrate in crates/pdfs-fuse/src/lib.rs.
Cause. The hydrate rebuilt the listing of every listed folder with a pass over all entries: 4,600 listed folders times 37,000 entries. That took 1.7 s on an idle machine (measured on a tree of that shape with entries of the real size), before the mount came up.
Fix. The listings are gathered in one pass over the entries (listed_children), which
takes 3 ms on the same tree.
Test: hydrated_listings_hold_each_folders_live_children in lib.rs: each listed folder
gets its live children, the root is not listed inside itself, and an unlisted folder gets no
listing.
B166 — A folder drops or empties a file whose create lands while it is listed
Status: Fixed.
Found: 2026-10-04, by the account run on the working tree after B165 (a wide directory, on
My files and on on-demand pairing 1): "concurrent unlinks left the wrong entries". Every
create and all 64 trashes landed as they should, so the drain was right and the listing was
wrong. The run logs no names. The simulation lost files that were never unlinked. The same
run read -starts-with-dash, 17 bytes, back as 0 bytes on on-demand pairing 1, just as its
create landed, so B163's fix does not cover it.
Where: list_children in crates/pdfs-fuse/src/lib.rs.
Cause. A folder whose listing was dropped is listed from Drive, then overlaid with what the queue still holds for it (B132). The queue was read after Drive answered. A create that landed in between was in neither: Drive had taken the listing before the node existed, and the queue had let go of it by the time it was read. The file was missing until the folder was listed again. On Wi-Fi, with 16 workers landing creates while the event feed drops the folder's listing, this happened in almost every run of the wide-directory case. When Drive's listing already had the node, but from before its create was committed, the node had no revision, and the file was listed at 0 bytes until its size was read again (the B163 state, reached through a listing instead of the read-back).
Fix. The queued children are read before Drive is asked as well as after. A node queued before and gone from the queue after has landed meanwhile, and is listed as the DB has it now, under its name. One unlinked or moved away meanwhile is no longer a child of the folder in the DB, so it stays out.
Test: half_of_a_wide_folder_unlinked_at_once_leaves_the_other_half in sim/daemon.rs
(failed 4 of 4 runs before the fix, passed 8 of 8 after), and
a_listing_taken_before_a_create_commits_keeps_the_landed_file, which holds Drive's answer to
the listing until the create has landed (the daemon listed the file at 0 bytes before the
fix).
B165 — A stop waits on a stuck thread until systemd kills the daemon
Status: Fixed (unverified).
Found: 2026-10-04, by the account run on 2.10.0 (cleanup after on-demand pairing 2). The
daemon was stopped at 15:53:04 and logged "control socket closing", but never "sync engine
stopping" or "daemon stopping". systemd killed it 15 s later (TimeoutStopSec). The run's
cleanup then found no control socket, because the new daemon had not opened it yet, and left
its folders behind.
Where: stop_workers in crates/pdfs-fuse/src/mount.rs; mount_once in
crates/pdfs-cli/src/main.rs.
Cause. A stop joined every background thread and waited as long as each one took. Every clean stop in the run logged the sync engine stopping just before the control socket closed. This one did not, so the engine was busy. It was most likely in a mirror pass over the run's mirror folder: that pass calls Drive with no deadline and does not check for a stop. The thread is inferred from the log, not observed.
Fix. The stop waits up to 10 s for the background threads and the FUSE workers together, then names any thread still running in a warning and exits. A queued op it cut off is sent again on the next start, as after a kill. The runtime is shut down with a 2 s limit, so a blocking task cannot hold the exit either. The pass itself still has no deadline (STATUS, "Any time").
Test: a_stop_names_the_workers_still_running_at_its_deadline in mount.rs.
B164 — A folder lists a new file twice while its create lands
Status: Fixed. Found: 2026-10-04, by the account run on 2.10.0 ("a wide directory", on-demand pairing 2). Eight threads wrote 128 files into a new folder, and the listing that followed held 129 entries. Each name was created on Drive once, and nothing was logged.
Where: overlay_queued in crates/pdfs-fuse/src/lib.rs.
Cause. A listing from Drive shows what is still queued for the folder over what Drive has (B132). A create Drive has made but whose answer is still on its way, or which the drain is still reading back, is in both: Drive lists it under its new uid, the queue under its stand-in. That listing only happens when the folder's listing was dropped, and the feed's event for a new file does that when it comes before the drain has noted the create as its own. Both were listed under the same name.
The run logged nothing for the folder, so the event is inferred, not observed. It is the one path found that lists a name twice.
Fix. A node queued in the folder holds its name there: Drive's node of the same name is left out of the listing until the queued op lands.
Test: simulation test a_create_drive_lists_before_it_is_read_back_is_listed_once (failed
before the fix: the folder listed c.bin twice).
B163 — A file read right after its create landed is empty
Status: Fixed.
Found: 2026-10-04, by the account run on 2.10.0 ("unusual but legal names", on-demand
pairing). The case wrote some twenty files into a new folder, listed it and read each back.
-starts-with-dash read 0 bytes for 17. Its create had landed with the bytes a moment before.
No revision was queued for it and nothing was logged.
Where: Core::read_back in crates/pdfs-fuse/src/lib.rs.
Cause. After a create, the drain reads the new node back and adopts that answer into the tree. Drive can answer a read before it lists a node it has just made, so an empty answer is asked again (B143). It can also list the node before the create's revision is committed, as a file with no active revision: no claimed size and nothing on storage, so its size reads as 0. The adoption took that answer, and the tree showed the new file as empty. The folder was made here, so its listing came from the tree. The size fetch the listing starts met the same answer and left the 0. The kernel capped the read at it.
The run logged nothing for the file, so the stale answer is inferred, not observed. It is the one path found that leaves a landed file empty with no op queued.
Fix. read_back treats a file without an active revision like a node Drive does not list
yet: it asks again after 250 ms and 1 s. If both answers are still without a revision, it
reports the node as not found, and the drain adopts it later, as it does for a node Drive
never lists in that time.
Test: simulation test a_file_read_back_before_drive_committed_it_keeps_its_bytes (failed
before the fix: the listing showed the file with 0 bytes). The fake Drive can now answer a
file's next reads as Drive lists a file that is not committed yet
(FakeClient::answer_reads_unrevised).
B162 — A folder removed and made again at once loses what goes into it
Status: Fixed.
Found: 2026-10-04, by the simulation's slow-link profile (one_client_on_a_slow_link,
seed 4). The client removed folder z and made z again, with z/x and z/x/y inside. All
three stayed in the mount but never reached Drive. The log showed the new z "adopting" the old
one.
Where: Core::adoptable_twin in crates/pdfs-fuse/src/drain.rs.
Cause. A create that finds its name taken looks for a node an earlier, unanswered attempt of its own made (B127), and adopts it. The old folder's trash was still on the wire, so Drive refused the new folder's name, and the listing showed the old folder, not in the tree any more and made a moment ago. The new folder took it for its own twin. What went into the new folder was created in the old one, and the trash then took it all along.
Fix. The twin search skips nodes this mount trashed (its hidden set), as listings already
do. The create then waits for the trash to free the name (name_is_held), or takes the name
at once if the trash has landed (B144).
Test: simulation test a_folder_made_again_while_the_old_one_is_trashed_is_a_new_folder
(failed before the fix: Drive ended up empty). The fake Drive can now hold a trash
(FakeClient::hold_next_trash).
B161 — A file moved while it was being created shows up under its old name
Status: Fixed.
Found: 2026-10-04, by the simulation's slow-link profile (one_client_on_a_slow_link,
seed 3). A file written and then moved at once ended up on Drive as y/z/d, but the mount
showed it as y/z/c.bin, its name before the move, and d was missing.
Where: Core::adopt_real_uid in crates/pdfs-fuse/src/drain.rs.
Cause. A move made while a file's create is on the wire waits in the queue until the create lands. Once the create is retired, the drain reads the new node back from Drive and adopts it. The tree keeps the name the user gave only while the move is still queued. The drain asked about that after the read-back. On a slow link the move could land in between: the answer still had the old name, and the move was no longer queued, so the old name won.
Fix. The drain asks whether a move is queued before it reads the node back. It still asks again under the tree's lock (B152).
Test: simulation test a_file_moved_while_its_create_was_on_the_wire_keeps_its_new_name
(failed before the fix: the new folder listed nothing). The fake Drive can now hold a read's
answer, taken when the read arrives, until the test releases it
(FakeClient::hold_answer_to_next_read).
B160 — A new folder reaches Drive only half a minute later
Status: Fixed.
Found: 2026-10-04, by the simulation's slow-link profile (one_client_on_a_slow_link,
seed 4). Folders made last before the client waited for its queue reached Drive 30 s late. The
log showed each mkdir queued, then claimed exactly 30 s later.
Where: Core::queue_local_node in crates/pdfs-fuse/src/lib.rs.
Cause. The drain workers sleep until an op is due, or for up to 30 s when none is. Every other way of queueing an op (a write, a rename, a trash) wakes them. Queueing a folder's create did not, so with the workers asleep it waited for their next poll. A file's create did not need the wake: its bytes ride on the create, and the release that attaches them wakes the drain.
Fix. Queueing a folder's create wakes the drain.
Test: simulation test a_folder_made_while_the_drain_is_idle_reaches_drive_at_once (failed
before the fix: the folder was still queued after 10 s).
B159 — A write right after a restart fails with "Resource temporarily unavailable"
Status: Fixed (unverified).
Found: 2026-10-04, by the simulation's flaky-link profile (one_client_on_a_flaky_link,
seed 4, one of ten runs on c2f9344). Half a second after the client restarted, a write into its
folder answered EAGAIN, and nothing was logged. The folder was still a queued create.
Where: Core::ensure_children in crates/pdfs-fuse/src/lib.rs.
Cause. A lookup in the root also places the "Shared with me" entry. It reads that entry's
plan from the database against a snapshot of the root listing, then publishes it only if the
listing has not changed meanwhile, and answers EAGAIN if it has. Right after a restart the
root is listed again while the first lookups come in, so a lookup could lose that race. Listing
a shared folder answers EAGAIN the same way when a newer refresh overtook it. No caller
retries, so the error reached the application.
Fix. ensure_children takes a listing again, up to four times, when an attempt answers
EAGAIN, and logs each retry at debug level.
Test: unit test a_listing_changed_while_published_is_taken_again. A stress test with
lookups and creates in the root at once did not hit the window (about 1,900 plans, none stale),
so the race itself has no regression test.
B158 — A file renamed twice in a row ends up under the first name, again
Status: Fixed (unverified).
Found: 2026-10-04, by the simulation's Wi-Fi profile (one_client_on_a_slow_link, seed 3,
four of ten runs on 4676a4e). A file was renamed to b.txt, whose name a create on the wire
still held, and then to e.md. Drive renamed it to e.md and then back to b.txt. The
liveness check found b.txt on Drive and e.md in the mount.
Where: Db::record_op_failure and Db::release_op_claim in
crates/pdfs-core/src/db/ops.rs.
Cause. B141's fix drops a rename that fails while a newer one is queued behind it. The worker records the failure first and releases its claim after that. The second rename came in between: the first was still claimed, so the second queued behind it, but the failure had already been recorded with nothing behind it, so the first was kept with a backoff. Once released, it was not due and the second was, so the second landed first. More workers made the gap easier to hit.
Fix. Releasing the claim of a rename drops it as well when a newer rename of the same node is queued behind it. Both checks match the node by local id as well as by uid.
Test: unit test a_rename_on_the_wire_lands_before_the_one_that_replaces_it, extended.
Seeds 2 and 3 of one_client_on_a_slow_link pass eight replays each.
B157 — A second write to a file whose upload was not read back fails with EIO
Status: Fixed (unverified).
Found: 2026-10-04, by the simulation's Wi-Fi profile (one_client_on_a_slow_link, seed 4,
every replay on 16fca10, none on dbc4d9a: sixteen drain workers changed the timing). After an
upload of c.bin landed, the read-back of the node went unanswered. A write at an offset
followed, and the next write to the file failed with EIO. The daemon logged "refusing write
over incomplete queued revision".
Where: Core::refresh_after_upload and Core::revision_conflict in
crates/pdfs-fuse/src/drain.rs.
Cause. The uploaded bytes are cached under the revision the read-back returns (B123). With no
read-back, the drain let go of the staged blob and cached nothing. A partial write opened next had
nothing to fill its gaps from, so it queued an incomplete revision, and a write open over one is
refused. Had it been filled, the write would have landed as a conflict copy of itself: the
revision id of the upload was never recorded as ours (own_sealed_rev).
Fix. When the read-back fails, the uploaded bytes are cached under the size and mtime the tree still has for the file, which is the base a write opened next takes, unless a newer write has taken over. The SHA-1 of those bytes is kept in memory, and a queued write that finds the remote holding them chains onto it instead of forking. After a restart the write still forks a conflict copy, which loses nothing.
Test: sim test a_file_whose_upload_was_not_read_back_takes_two_writes_at_once, with the
fake Drive's new lose_read_back_of_next_revision. Without the fix the second write fails with
EIO; with only the cache fix, it lands as a conflict copy.
B156 — A file written again after its create lost its answer lands as a conflict copy
Status: Fixed (unverified).
Found: 2026-10-04, by the simulation's Wi-Fi profile (one_client_on_a_slow_link, seed 4,
two of six replays on e5f3f61). The create of a.txt timed out twice, but Drive had made the file
with the bytes of the first write. A later write replaced the blob the queued create carried. The
third attempt found the name taken and created a (sync-conflict …).txt; a.txt kept the old
bytes.
Where: Core::adoptable_twin and Core::withdrawn_twin in crates/pdfs-fuse/src/drain.rs.
Cause. A twin is adopted when it holds the bytes the create uploads (B127), and only the blob the op held at the retry was hashed. The twin held the bytes an earlier attempt sent.
Fix. Each attempt records the SHA-1 of the blob it uploads, by op id, in memory. A twin holding
any of them is ours. One holding older bytes than the op's is adopted, and the op's blob goes up
as a revision of it. Kept on the op's row since B183, so a restart no longer forgets them. Sim test a_file_written_again_after_its_create_lost_its_answer_lands_once; the fake
Drive's held and lost create answers now combine.
B155 — A write whose upload lost its answer lands as a conflict copy of itself
Status: Fixed (unverified).
Found: 2026-10-04, by the simulation's Wi-Fi profile (one_client_on_a_slow_link, seed 2,
second replay on 7269cb5). x/z/y/a.txt was written after its create landed. The revision upload
reached Drive, but the answer was lost and the call timed out. On the retry, Drive held
a (sync-conflict …).txt beside the file, with the same bytes.
Where: Core::revision_conflict in crates/pdfs-fuse/src/drain.rs.
Cause. A queued write remembers the revision it was made against. The retry found the file one
revision further, sealed by the lost upload, which this daemon never heard of, so it was not
recorded as our own (own_sealed_rev). The write was taken for a change from elsewhere and kept as
a conflict copy.
Fix. A complete write whose bytes the remote already holds, by size and SHA-1, is not a
conflict: it is sent again as a revision of the same bytes. Sim test
a_write_whose_answer_was_lost_lands_once, with the fake Drive's new
lose_reply_to_next_revision.
B154 — A file made under the name of a create on the wire lands as a conflict copy
Status: Fixed (unverified).
Found: 2026-10-04, by CI on 7269cb5 (one_client_on_a_good_link, seed 9, every run). c.bin
was renamed to d while its create was on the wire, and a new c.bin was made. The new file landed
as c (sync-conflict …).bin, the mount showed that name, and unlink c.bin answered ENOENT.
Where: the failure path of Core::run_pending_drain in crates/pdfs-fuse/src/drain.rs.
Cause. A regression from the B153 fix. The new file's create found c.bin held by the first
create, which was still on the wire, and failed as a wait. The B153 fix then woke the ops waiting
for the name. The database cannot see a create on the wire, only queued trashes and renames, so
it found the name free and the op was due again at once. Four waits went in 70 ms, and the create
took a conflict name.
Fix. No wake while another create on the wire holds the name. It frees the name when it
lands, or hands it to the rename queued behind it, which wakes the waiters when it lands. Sim
test a_file_made_under_the_name_of_a_create_on_the_wire_waits_for_it, with the fake Drive's
new hold_reply_to_next_create.
B153 — An op waiting for a name sat out its backoff after the name was free
Status: Fixed (unverified).
Found: 2026-10-04, by the simulation's LAN profile (one_client_on_a_good_link, seed 4,
about one run in five, also before the B151 change). d was renamed over c.bin while the
create of the old c.bin was on the wire. The queues did not drain within 4 s: the rename still
owed c.bin is still held by a node whose trash or rename is queued.
Where: the failure path of Core::run_pending_drain in crates/pdfs-fuse/src/drain.rs.
Cause. The rename found the name held by the queued trash of the old c.bin and failed. The
trash landed a millisecond later and woke the ops waiting for the name
(Db::wake_ops_waiting_for), before the rename had recorded its failure, so there was nothing to
wake. The failure then set the doubled backoff of a network error, 4 s on the first attempt and
more after it.
Fix. A held name fails as NameHeld, which backs off by the minimum (2 s) without doubling,
and the op wakes the waiters for its name again once its failure is recorded. Unit test
a_held_name_is_a_wait_and_other_failures_are_not; the seed is the regression check.
B152 — A rename made as its create landed was undone in the tree
Status: Fixed (unverified).
Found: 2026-10-04, by the simulation's LAN profile (one_client_on_a_good_link, seed 4,
about one run in ten). a.txt was renamed to b.txt 3 ms before its create finished landing.
The next rename, b.txt to e.md, answered ENOENT, and the node row kept a.txt although
Drive had b.txt.
Where: Core::adopt_real_uid in crates/pdfs-fuse/src/drain.rs.
Cause. Adoption puts the node Drive made into the tree, but keeps the tree's name and folder while a rename is queued. It asked whether one was queued before it wrote the remote uid into the queue and took the tree lock. A rename between the two was overwritten with the name the create was sent with, in the tree and in the node row, which was written after the lock.
Fix. Adoption asks again under the tree lock. Core::queue_rename queues its op before it
renames the entry under that lock, so a renamed entry always has its op. The node row is written
through the lock's outbox, in order with the rename's own write. No test: the window is inside
one call and no fake hook reaches it; the seed is the regression check.
B151 — A file deleted after its create lost its answer stays on Drive
Status: Fixed (unverified).
Found: 2026-10-04, by the simulation's Wi-Fi profile in CI (one_client_on_a_slow_link,
seed 2). x/a.txt was made, written and deleted within a third of a second. The queues never
drained, and Drive kept x/a.txt with the bytes of its create.
Where: drop_doomed_ops in crates/pdfs-core/src/db/ops.rs, and Core::drain_trash in
crates/pdfs-fuse/src/drain.rs.
Cause. A create that times out may still have made the file; only its answer was lost (as in B127). The create stays queued with a backoff. A delete before the retry dropped the queued create as if nothing had been sent, so the node Drive had made kept the deleted file's name and bytes. Only a create on the wire at the time of the delete was turned into a trash (B129), and when that create failed, the trash took it as never made.
Fix. A create that was sent, on the wire or before, becomes a trash that keeps the create's
folder and blob. The drain looks in that folder for a node under the create's name that the
create would have adopted (B127): made around when it was queued, and empty or holding the
blob's bytes. It trashes that node, and then discards the blob. A new create of the same name in
that folder waits for the trash. Sim test
a_file_deleted_after_its_create_lost_its_answer_leaves_drive.
B150 — A conflict copy whose folder is gone retries forever
Status: Fixed (unverified).
Found: 2026-10-04, by the journal check of the account run on a6f36d5: 33 warnings, one
every two minutes for the whole run, all "queued write conflicts; keeping a conflict copy" for a
write.bin from an earlier run's B79 case, each followed by DoesNotExist from Drive.
Where: Core::keep_as_conflict_copy in crates/pdfs-fuse/src/drain.rs.
Cause. A queued write to a file that was trashed or deleted on Drive is kept as a conflict copy beside the file. When the file's folder was deleted too, the upload into it failed, the op backed off and tried again, with no end. The copy only fell back to the root when the tree had forgotten the file, not when Drive had lost its folder.
Fix. The copy checks that the folder is still on Drive and not trashed, and lands in the
location's root otherwise. Sim test a_queued_write_whose_folder_was_trashed_lands_in_the_root.
Left open. How the B79 case's bytes stayed in staging/ after the file was deleted, so that
the next start queued them again ("reconciled orphaned staged writes queued=1"), is not known yet.
B149 — pdfs rename or pdfs move right after a rename through the mount is undone
Status: Fixed (unverified).
Found: 2026-10-04, by the account run on a6f36d5: the B121 case failed on every on-demand
mount. After renaming and moving a file through the mount, it ran pdfs rename and pdfs move,
and pdfs ls then showed the file in the right folder under the name the mount had given it.
Where: Core::rename and Core::move_to in crates/pdfs-fuse/src/lib.rs.
Cause. Since changes made through the mount are queued (local-first), the mount's rename was
still in the queue when the CLI's requests came. Those went straight to Drive, landed first, and
the queued rename landed after them and put its own name back. A file whose create was still
queued could not be renamed by pdfs at all, since Drive does not know it yet.
Fix. pdfs rename and pdfs move within a mount are queued whenever the mount would queue
the same change, or the node has anything queued. The tree and the kernel are updated at once,
as for a rename through the mount, and a taken name is refused up front. A node whose create is
still queued has that create's target rewritten instead. A move between two locations still goes
straight to Drive.
B148 — A folder pinned before it has reached Drive loses its pin
Status: Fixed (unverified).
Found: 2026-10-04, while planning the move of queued ops to local ids: every table keyed by
uid was checked for what happens when a placeholder lands, and pins was the one no landing
touched.
Where: adopt_placeholder_row_tx in crates/pdfs-core/src/db/nodes.rs.
Cause. A folder made through the mount has a placeholder uid (local~…) until its mkdir lands.
Pinning it then stored the placeholder in pins.uid. The landing moved the node row, the children
and the queued ops to the real uid, but not the pin, which then named no node: the folder was no
longer kept offline and the pin list showed an entry for nothing.
Fix. The landing moves the pin to the real uid in the same transaction. A pin the real uid already has wins, and the placeholder's is dropped.
B147 — A new file made under a name just renamed away can share the first file's node
Status: Fixed (unverified).
Found: 2026-10-04, by sim::run::tests::three_clients_one_writer_each, seed 2, in about one
run of eight. Client 1 made a.txt, renamed it to b.txt, and made a.txt again. Drive ended
up without b.txt, and e.md, a later rename of the new a.txt, held b.txt's bytes.
Where: Core::adoptable_twin in crates/pdfs-fuse/src/drain.rs.
Cause. The first file's create went out as a.txt and the rename was queued behind it. While
that create was on the wire, the second file's create found a.txt taken by a fresh, empty file,
which is how the node of an unanswered create of its own looks, and adopted it. Both local files
landed on the one node, and the first file's rename took it to b.txt.
Fix. A create no longer adopts a node another create is making under the same name and folder: it waits for that one to land, as for any name a queued change holds. Nor does it adopt a node already in the tree, which was answered for. An unanswered twin the tree has already listed lands as a conflict copy, which loses nothing.
B146 — A file renamed just as its upload lands can go back to its old name
Status: Fixed (unverified).
Found: 2026-10-04, by the CI simulation run (sim::run, profile one-client-echoes, seed 4).
Client 0 renamed a.txt to b.txt, and a millisecond later renaming b.txt to e.md failed
with ENOENT. Not reproduced locally.
Where: Core::refresh_after_upload in crates/pdfs-fuse/src/drain.rs.
Cause. After an upload lands, the drain reads the node back and puts Drive's metadata in the tree. It keeps the tree's name only while a rename is queued, and it asked the queue before taking the tree's lock. A rename queued between the two had already renamed the entry, and Drive's metadata, which still had the old name, put it back.
Fix. The drain notes the entry's name and folder before it asks the queue. If either changed by the time it holds the lock, a rename came in between, and the tree keeps it.
Test: none. The window is between two statements and is not hit on purpose.
B145 — Renaming a just-made download to its final name can fail with "Device or resource busy"
Status: Fixed (unverified).
Found: 2026-10-04, by the account acceptance run (B70 check). The browser-style rename
b70-download.zip.crdownload → b70-download.zip failed with EBUSY. The journal shows
swept parked creates released=0 dropped=1 3 ms after the create was queued, then
queued create vanished under a rename.
Where: Core::sweep_parked_creates and park_verdict in crates/pdfs-fuse/src/drain.rs.
Cause. A transient file's create is parked until the rename that gives it its final name.
queue_local_node queues the create before the caller writes the node's row. The park sweep
reads a parked create with no row as a file that was deleted and drops it. A sweep that ran
between the two steps dropped a live file's create, and the rename found nothing to rename.
Fix. A parked create with no row is dropped only once it is a minute old and not open. Before that it is kept, like any other park.
Test: the unit test a_parked_create_is_dropped_only_once_its_file_is_gone_and_closed. The
window is a few milliseconds wide and is not hit on purpose.
B144 — A rename can land under a conflict name although its target was freed
Status: Fixed (unverified).
Found: 2026-10-04, by the simulation runs (sim::run, profile one-client-echoes, seed 4).
Client 0 renamed d to c.bin while the trash of the old c.bin was queued. The rename landed
as c (sync-conflict …).bin, and the client's tree and Drive disagreed.
Where: Core::drain_rename_in_place, Core::drain_rename and Core::create_local_node in
crates/pdfs-fuse/src/drain.rs.
Cause. When Drive answers that the name is taken, the drain asks whether a queued trash or rename still holds it, and if so waits for that to land. The trash landed between Drive's answer and the check. Nothing held the name any more, so the drain took the answer as a real clash and fell back to a conflict name.
Fix. When nothing holds the name, the drain sends the request once more under the wanted name before falling back to a conflict name.
Test: seed 4 of one-client-echoes passes repeatedly. The window is too narrow to hit on purpose.
B143 — A write to a file whose create landed as the link went down lands as a conflict copy
Status: Fixed (unverified).
Found: 2026-10-04, by the simulation runs (sim::run, profile three-clients, seed 1).
Client 0 made y/z/b.txt while the link was down. When the link came back, the create landed and
the link went down again right after it. A write to the file was then kept as
b (sync-conflict …).txt: "expected 56 bytes at mtime 1791068288, found 56 at 1791068289". Only
client 0 writes that file.
Where: Core::drain_local_node, Core::drain_op and Core::finish_adoption in
crates/pdfs-fuse/src/drain.rs.
Cause. The gap B131 closed, for a link that drops instead of a lookup that is late.
Core::read_back asks again only while Drive answers that it does not list the node yet. When
the link is down it gives up, and adopt_real_uid fails after the create's op has retired. The
tree keeps the placeholder's node under the real uid, and nothing reads it back later. A write
opened after that takes the local clock's mtime as its base. Drive stamped the create with the
second it arrived, so the drain saw a revision it did not know and kept the client's own write as
a conflict copy.
Fix. A create whose read-back fails is remembered, under the uid Drive gave it. Before the
drain sends the next op for that uid, it reads the node back and rebases handles and queued writes
onto it, as adopt_real_uid would have done. If the read-back fails again, the op is retried
later, as for any other failed request.
Test: the daemon test a_write_over_a_create_whose_read_back_failed_is_not_a_conflict. The
fake Drive takes the link down as soon as it has answered the create. Without the fix, the test
fails on the conflict copy.
B142 — Renaming a new file as its upload lands can fail with "Device or resource busy"
Status: Fixed (unverified).
Found: 2026-10-04, by the simulation runs (sim::run, profile three-clients, seed 3).
Client 1 renamed c.bin, a file whose create was still queued, to f.txt, and got EBUSY.
Where: serve_rename in crates/pdfs-fuse/src/filesystem.rs.
Cause. A file whose create is queued is renamed by rewriting the name in the create op. When
the create landed between the lookup and the rewrite, the op was gone, and the rename answered
EBUSY as if the create had vanished.
Fix. When the op is gone, the rename looks up the node's row. If the create has landed, the row already holds the uid Drive gave the file, and the rename is queued against that uid, as for any other file.
Test: seed 3 of three-clients passes repeatedly. The window is too narrow to hit on purpose.
B141 — A file renamed twice in a row can end up under the first name
Status: Fixed (unverified).
Found: 2026-10-04, by the simulation runs (sim::run, profile one-client-echoes, seed 2).
e.md was moved into x, and right after that into the client folder. Drive kept it in x,
and the liveness check found it in a different place on Drive than in the mount. In about one of
three replays.
Where: Db::enqueue_op and Db::claim_next_due_op in crates/pdfs-core/src/db/ops.rs.
Cause. A rename replaces the rename queued before it, by deleting that row. When a drain worker had already taken the older row, deleting it also lifted the rule that one node has one op in flight. A second worker sent the newer rename at once, while the first was still retrying its own after Drive called it out of date (B111). The older one landed last.
Fix. A rename a worker holds stays queued, and the newer one waits behind it. If the older one fails, it is dropped instead of being retried after the newer one.
Test: the unit test a_rename_on_the_wire_lands_before_the_one_that_replaces_it. Seed 2 of
one-client-echoes passes repeatedly.
B140 — Files in a folder that lands as the daemon stops are gone after a restart
Status: Fixed (unverified).
Found: 2026-10-04, by the simulation runs (sim::run, profile one-client-flaky, seed 4),
replaying B132. The client folder c0 was made with files in it, and its mkdir landed after the
mount had stopped. Right after the restart, a write to c0/a.txt failed with ENOENT, and the
listing of c0 was empty. In about half of the replays.
Where: adopt_placeholder_row_tx in crates/pdfs-core/src/db/nodes.rs and Core::hydrate in
crates/pdfs-fuse/src/lib.rs.
Cause. When a folder lands, the rows below it are handed to the real uid by their
parent_uid column only. The node each row holds (node_json) still named the placeholder. A
running daemon rewrites the rows from its tree on the next change, so this did not show. A restart
builds the tree from node_json, found no row for the placeholder and left the children out of the
folder.
Fix. The landing rewrites the parent in node_json as well, in the same statement.
Test: the unit test a_landed_folder_hands_its_children_over_in_the_stored_node_too. Seed 4
of one-client-flaky passes ten replays in a row.
B139 — A file closed while its first upload lands keeps its last write in staging
Status: Fixed (unverified).
Found: 2026-10-03, by the simulation runs (sim::run, profile three-clients, seed 1), at the
end check for lost data. A file made through the mount was written and renamed while its create
was queued. Its handle was closed just as the create landed. The daemon logged "cannot queue
write; bytes kept in staging" with "access changed before the write could be queued", and Drive
kept the bytes from before the last write.
Where: Core::queue_revision in crates/pdfs-fuse/src/lib.rs and release in
crates/pdfs-fuse/src/filesystem.rs.
Cause. A write handle keeps the uid the node had when it was opened. When the create lands,
finish_create retires the op and the placeholder row becomes the real node's, before the tree
moves the open handle over. A handle closed in that gap still named the placeholder, which no
longer had a row, so the access check refused it and the bytes went to staging instead of the
queue.
Fix. On close, a handle that names a placeholder looks up the node by its inode's local id, which survives the landing, and queues the write against the real uid when the node has one.
Test: the unit tests a_landed_placeholder_keeps_its_lid and
an_inode_gives_back_the_lid_it_was_made_from. Seed 1 of three-clients passes.
B138 — A file created where another is leaving lands as a conflict copy
Status: Fixed (unverified).
Found: 2026-10-03, by the simulation runs (sim::run, profile one-client-lan, seed 9).
c.bin was created and renamed to d while its create was on the wire, and a new c.bin was
made at once. The new file landed on Drive under a conflict name.
Where: drain_local_node, drain_rename and drain_rename_in_place in
crates/pdfs-fuse/src/drain.rs.
Cause. When Drive refuses a name, the drain waits only if a queued trash holds it
(has_pending_trash_named); otherwise it lands the op under a conflict name. A create on the wire
holds the name it was sent with until the rename queued behind it lands, and a queued rename away
from a name holds it until it lands. Several workers drain at once, so the new create could reach
Drive first.
Fix. The drain also waits while another create on the wire was sent with that name in that folder, or a queued rename moves a node away from it. Those waits end after a few attempts, as two files swapping names each hold the name the other wants. When a rename or trash frees a name, the ops that failed on it are made due at once, and so is a create renamed away from a name it failed on, whether the rename came before the failure was recorded or after.
Test: the unit test a_queued_rename_away_from_a_name_holds_it_until_it_lands. Seed 9 of
one-client-lan passes.
B137 — A queued change can be deleted by an older one landing
Status: Fixed (unverified).
Found: 2026-10-03, by the simulation runs (sim::run, profile one-client-lan, seed 9), at the
end check for lost data. A file renamed and then deleted stayed on Drive, and the queue was
empty.
Where: the pending_op table (crates/pdfs-core/src/db/migrations.rs) and Db::delete_op
in crates/pdfs-core/src/db/ops.rs.
Cause. pending_op.id was a plain INTEGER PRIMARY KEY, so SQLite hands out the largest id
plus one, and an id comes back once the newest row is deleted. Deleting a node drops its queued
ops and queues a trash in one transaction, so the trash could take the id of the rename the drain
was sending. When that rename landed, the drain deleted its row by id, which was now the trash.
Fix. Schema 37 rebuilds pending_op with AUTOINCREMENT, so an id is never given out
again.
Test: the unit test a_dropped_op_id_is_never_given_out_again.
B136 — A file renamed while its first upload is on the wire goes back to its old name
Status: Fixed (unverified).
Found: 2026-10-03, by the simulation runs (sim::run, profile one-client-flaky, seed 3), at
the end check for lost data. A file made offline was renamed after the drain had taken its create.
The create landed under the old name and the new name was gone. Every replay of the seed.
Where: Db::finish_create in crates/pdfs-core/src/db/ops.rs, drain_local_node and
adopt_real_uid in crates/pdfs-fuse/src/drain.rs.
Cause. A rename of a node whose create is still queued rewrites the create's parent and name
(Db::rewrite_op_target) instead of queuing a rename. The drain copies the op when it takes it,
so a rewrite while the create was on the wire changed only the row. When the create landed,
finish_create deleted the row without looking at it again, and adopt_real_uid put the node
back where Drive had it. The move was lost on Drive and in the tree.
The same seed showed a second fault. When a folder made offline landed, the nodes in it still named the folder's placeholder uid as their parent in memory. The next write of one of them put the placeholder back on its row.
Fix. The drain passes where it sent the create and where it landed (a conflict name, or the
root when the folder was gone) to finish_create. When the row names another parent or name than
it was sent with, the same transaction queues a rename of the real uid to it, with the landed
parent and name as the original ones. adopt_real_uid keeps the tree's name and parent while
that rename is queued. When a node lands, its children's parent is changed to its real uid.
Test: the unit test a_rename_made_while_its_create_uploads_is_queued_for_the_landed_node.
Seed 3 of one-client-flaky passes.
B135 — A file just moved is sometimes not found by its new path
Status: Fixed (unverified).
Found: 2026-10-03, by the simulation runs (sim::run, profile three-clients, seed 1), once
B131 was fixed. rename a.txt -> y/f.txt succeeded, and the rename y/f.txt -> y/a.txt two
milliseconds later failed with ENOENT. Once in about four replays of the seed.
Where: the online path of serve_rename in crates/pdfs-fuse/src/filesystem.rs.
Cause. After the move lands on Drive, serve_rename settles the tree with State::relocate.
It took the inode lock directly (state.lock()) instead of through Core::state, whose guard
writes what a mutation owes the database once the lock is released (DB1). So the writes stayed in
the state's outbox: both folders no longer listed, and the node's new parent. Whichever thread
next released a guard applied them, after it had let go of the lock. The lookup of y/f.txt
found y unlisted in memory and asked the database, which still marked y as listed and placed
the file in the root, so it rebuilt y without the file. It took another thread taking the lock
in between, so it happened only now and then.
Fix. serve_rename relocates through Core::state, so the database has the move before the
reply.
Test: the unit test an_online_rename_writes_the_move_through_before_it_answers. With the
fix, seed 1 of three-clients passes repeatedly with PDFS_SIM_KNOWN=fail.
B134 — A file deleted online while the drain checks its write is kept as a conflict copy
Status: Open, with local_first off only. By default an unlink drops the write and queues
the trash in one step, and the drain sends the trash.
Found: 2026-10-03, by the simulation runs (sim::run, profile three-clients, seed 3).
c.bin was written, renamed to f.txt and deleted, all online. The drain logged "queued write
conflicts; keeping a conflict copy" for f.txt, with "the file was trashed remotely". The copy
was named after the file, not recovered-…, so the tree still placed the file when the drain
looked. It did not come up again in five replays with debug logs.
Where: trash_child in crates/pdfs-fuse/src/filesystem.rs, and Core::drain_revision in
crates/pdfs-fuse/src/drain.rs.
Cause. An online unlink trashes the node on Drive first, and only then forgets it in the
tree and drops its queued write (discard_queued_ops, or withdraw_queued_ops for an open
file). A drain worker that has claimed the write and reads the node in between finds it trashed.
B105's check asks whether the op is still queued, and it still is, so the drain keeps the write
as a conflict copy of a file the user deleted. When the copy runs before the unlink discards the
staged blob, it lands next to where the file was.
Test: none. The simulation runs mount with local_first on, so they no longer reach it.
B133 — A file deleted while its upload is in flight can still come back as a recovered copy
Status: Fixed.
Found: 2026-10-03, by the simulation runs (sim::run, profile one-client-deletes, seed 3).
y/e.md was written, and the drain picked the write up. The step held the drain's read of the
node, deleted the file, and released the read. The next step wrote a.txt. The drain then logged
"queued write conflicts; keeping a conflict copy" for y/e.md as recovered-L4, with "the file
was trashed remotely". It is B105 again, though B105's fix was in place. Here the copy failed
because the blob was already gone.
Where: Core::drain_revision in crates/pdfs-fuse/src/drain.rs, Db::op_exists in
crates/pdfs-core/src/db/ops.rs, and the pending_op table in
crates/pdfs-core/src/db/migrations.rs.
Cause. B105's fix asks whether the op the worker holds is still queued, by id. pending_op.id
is an INTEGER PRIMARY KEY without AUTOINCREMENT, so SQLite gives a new row the largest id
in the table plus one, or 1 when the table is empty. The delete dropped the newest row, so the
next op queued took the same id. op_exists found that other op and said the write was still
wanted. A SQLite workload queues a write to the database right after deleting its journal, so
this is the usual case there, not a rare one.
Fix. With B137's: schema 37 gives pending_op ids with AUTOINCREMENT, so an op queued
after a delete never takes the dropped op's id.
Test: the simulation runs make the file again, or not, while the drain's read is held, and
fail on any conflict copy the drain keeps of it. They counted it instead until the run on
812eb3d passed every profile on twelve seeds without the count.
B132 — A folder whose mkdir is still queued is gone after a restart
Status: Fixed.
Found: 2026-10-03, by the simulation runs (sim::run, profile one-client-flaky, seed 4).
The first listing of the Drive root failed with the link down, and the client's folder c0 was
made offline. The link came back, and the client restarted with 8 ops still queued, c0's
mkdir among them. Right after the restart, a write to c0/a.txt failed with ENOENT. The kernel's
lookup of c0 in the root had answered ENOENT. A moment later the mkdir landed.
Where: Core::ensure_children in crates/pdfs-fuse/src/lib.rs, with
Core::relist_offline_listings and State::invalidate_listing.
Cause. A listing served offline is invalidated when the link comes back
(relist_offline_listings), which also clears its listed row. The running daemon does not
notice: the kernel still holds the dentries. After a restart, the first lookup in the folder
lists it on Drive. Drive does not hold a node whose create is still queued, or does not list a
fresh one yet (the lag of B118). ensure_children takes Drive's listing as the whole folder and
records it as complete, without the queued node's placeholder. Every path through that node fails
with ENOENT until something lists the folder again. The same may happen without a restart
whenever a folder holding a queued create is listed on Drive.
Fix. A listing read from Drive is laid over with what the queue holds for the folder
(Db::queued_children): a node a queued create, mkdir or rename put there is listed as the DB has
it, and one a queued rename moved away is left out. Replays of the seed after this fix turned up a
second cause, B140.
Test: the unit test queued_children_are_what_drive_does_not_list_yet. Seed 4 of
one-client-flaky passes ten replays in a row.
B131 — A write to a file whose create could not be read back lands as a conflict copy
Status: Fixed (unverified).
Found: 2026-10-03, by the simulation runs (sim::run, profile three-clients, seed 1).
a.txt was created while the link was down, and the link came back right after. The create
landed, and the drain logged "pending upload failed; will retry" for its placeholder with
"fetch node: no such file or folder". Half a second later a write to a.txt queued a revision.
The drain kept it as a (sync-conflict …).txt: "expected 19 bytes at mtime 1791005792, found
19 at 1791005793". No other device had touched the file.
Where: Core::drain_local_node and Core::adopt_real_uid in crates/pdfs-fuse/src/drain.rs.
Cause. drain_local_node retires the create's op (retire_create) before
adopt_real_uid reads the new node back. Drive can answer a create before a lookup finds the
node (B118), and then the read-back fails. By then the op is gone, so the "will retry" never
happens. The tree's entry has the real uid but still holds the placeholder's node: the local
clock's mtime and no revision id. rebase_open_writes and rebaseline_pending never run. A
write opened afterwards takes that node as its base. Drive stamped the file when the create
landed, and if that was in a later second than the placeholder's mtime, the drain sees the
remote revision changed and keeps the client's own write as a conflict copy.
Once the read-back succeeded, the same seed found a second gap. A write handle open on the file
is repointed at the real uid when the op retires, and its release fills its gaps from the content
cache. adopt_real_uid cached the create's bytes only after the read-back, which waits for Drive
to list the node. A release in between found nothing to fill from and queued an incomplete
revision, and the next write open was refused with EIO (as in B123).
Fix. adopt_real_uid reads the node back with Core::read_back, which asks again for a moment
while Drive does not list it yet, as the online create does (B118). Before the op retires,
Core::cache_created caches the uploaded bytes under the real uid, keyed by the size and mtime
the tree has for the file, which is what an open handle has as its base. adopt_real_uid caches
them again, keyed as Drive has the node, before it rebases anything onto it.
Test: seed 1 of three-clients passes with PDFS_SIM_KNOWN=fail. Before the fix it failed on
the conflict copy when the create landed in a later second than it was made, which depends on
timing, and then on the refused write.
B130 — A read of a file that a remote change invalidates can hang the daemon
Status: Fixed.
Found: 2026-10-03, by the simulation runs (sim::run, profile one-client-flaky, seed 1).
After the last step, a read of c0/z/f.txt never returned. The watchdog's stacks showed the
read's worker in read_range_remote, waiting in block_on for its fetch, and a runtime worker
running run_event_sync blocked in the kernel in inval_inode, waiting for a page lock. The
read had started at 05:13:38. Its 120 s timeout, a 20 s bounded call and every other timer
of the runtime all fired at 05:17:24, the moment the watchdog aborted the mount.
Where: apply_event in crates/pdfs-fuse/src/background.rs, which runs on the runtime and
flushes its invalidations through Core::for_each_mount and NotifyBatch::flush in
crates/pdfs-fuse/src/lib.rs, and Core::read_range_remote in crates/pdfs-fuse/src/reads.rs.
Cause. The kernel holds the page lock of the pages a FUSE read fills until the daemon
answers. inval_inode on the same file waits for that lock, so it waits as long as the read.
The event sync sends it from a runtime worker, which blocks that worker in the kernel. If that
worker was the one parked on the runtime's driver, nothing drives the timers or the I/O any
more: tokio hands the driver on only when a parked worker is woken for new work, and the others
sleep on a condvar. The read's fetch waits on a timer (here the fake's latency, on Drive an
HTTP response), so it never finishes and never lets go of the page lock. The daemon stays stuck
until some other work wakes a worker, and on an idle machine that may never happen.
Fix. Every notice to the kernel goes through off_the_runtime, which on a runtime worker
sends it from inside block_in_place: the worker's tasks, and the driver with them, go to
another thread while it waits. Anywhere else the notice is sent as before.
Test: a_kernel_notice_sent_from_the_runtime_leaves_its_timers_running in
crates/pdfs-fuse/src/lib.rs: the one worker of a runtime sends a notice that returns only once
a timer of that runtime has fired. The simulation runs still recognise it in the watchdog's
stacks (a tokio-rt-worker in state D inside NotifyBatch::flush); with
PDFS_SIM_KNOWN=fail, seed 1 of one-client-flaky fails on it, though it no longer reaches it
since phase 3.
B129 — A file deleted or replaced while its create is on the wire stays on Drive
Status: Fixed.
Found: 2026-10-03, by the simulation runs (sim::run, profile one-client-lan, seed 4).
a.txt was created, renamed to b.txt and then to e.md while its create was still queued.
A moment later rename f.txt -> e.md failed with EIO, and the daemon logged "rename failed"
with Drive's AlreadyExists.
Where: Core::remove_replaced and Core::discard_queued_ops in crates/pdfs-fuse/src/lib.rs,
trash_child in crates/pdfs-fuse/src/filesystem.rs, and drain_local_node and
upload_created_file in crates/pdfs-fuse/src/drain.rs.
Cause. A node whose create is still queued has no node on Drive, so deleting it or renaming
another file over it only drops the queued op (discard_queued_ops). That cannot stop a drain
worker that has already claimed the op. cancel_upload only reaches revisions, and the upload
of a create has no cancel flag. The create lands anyway. finish_create finds no op row to
retire, and the tree has already forgotten the local uid, so nothing trashes the new node.
Drive keeps a file that the mount no longer shows. Its name stays taken: an online rename onto
it, as here, fails with EIO, and a later create with that name lands as a conflict copy.
Fix. A delete or replacing rename no longer drops a create the drain has taken. It turns
the row into a trash of the placeholder, still claimed by the upload, which holds the name as any
queued trash does. When the create lands, finish_create points that trash at the landed node in
the same transaction, and the drain sends it at once; if that fails it stays queued with the usual
backoff. A create whose row is gone altogether gets such a trash too. A trash of a placeholder
whose create never landed is done without asking Drive.
Test: the simulation runs fail on it. They counted it instead until the run on 812eb3d
passed every profile on twelve seeds without the count.
B128 — A file created offline over one deleted offline can be lost
Status: Fixed.
Found: 2026-10-03, by the simulation runs (sim::run, profile three-clients, seed 3).
While the link was down, f.txt was deleted and a new f.txt was made. Its first version had
not landed yet, so Drive held it empty. When the link came back, the drain logged "the name is
held by our own unanswered create; adopting it" with the old file's uid, then "pending trash
landed" for that uid. The queue drained empty, and Drive held no f.txt, only the trashed old
one.
Where: drain_local_node and adoptable_twin in crates/pdfs-fuse/src/drain.rs.
Cause. Several workers drain at once, so the new file's create can run before the trash of
the old one has landed, and Drive refuses the name. On that refusal the drain first looks for a
node its own unanswered create made: one with that name, made within ADOPT_WINDOW_MS, and
empty for a file. The old file matched, because it was made a moment earlier and its bytes were
still queued. The drain adopted it, so the new bytes went to it as a revision, and the queued
trash then trashed it. Only after the adoption check does the drain ask whether a queued trash
holds the name (has_pending_trash_named), and the adoption does not exclude a node that has
one.
Fix. With B162's: the adoption skips a node this mount trashed (its hidden set), which a
queued trash puts it in, and the create waits for the trash to free the name (B144).
Test: the simulation runs fail on it. They counted it instead until the run on 812eb3d
passed every profile on twelve seeds without the count.
B127 — A file created offline lands twice when the answer to its upload is lost
Status: Fixed.
Found: 2026-10-03, by the simulation runs (sim::run, profile one-client-flaky, seed 1).
The upload of a queued create timed out after Drive had made the file. On the next attempt the
drain logged "name is taken remotely; creating under a conflict name". Drive ended with the file
and with a conflict copy of the same bytes next to it.
Where: drain_local_node and adoptable in crates/pdfs-fuse/src/drain.rs.
Cause. When a create finds its name taken, the drain adopts the node only if it looks like
one its own unanswered create made (adoptable_twin). It must be made within ADOPT_WINDOW_MS
and, for a file, be empty. A queued create uploads its bytes in the same call, though
(upload_created_file). So the node a lost answer leaves behind holds those bytes, is not
adopted, and the file is uploaded a second time under a conflict name.
Fix. A file is also adopted when Drive's SHA-1 of its content is that of the op's blob. The
blob then counts as uploaded. Sim test a_create_whose_answer_was_lost_lands_once. A write that
replaced the blob after the lost upload still forks a copy: the node holds the older bytes.
Test: the simulation runs fail on it. They counted it instead until the run on 812eb3d
passed every profile on twelve seeds without the count.
B126 — A file created offline under a name an offline rename freed lands as a conflict copy
Status: Fixed.
Found: 2026-10-03, by the simulation runs (sim::run, profile three-clients, seed 3).
While the link was down, b.txt was renamed to c.bin, and a new b.txt was made. When the link
came back, the drain logged "name is taken remotely; creating under a conflict name" for
b.txt. The new file ended up only as b (sync-conflict …).txt.
Where: drain_local_node in crates/pdfs-fuse/src/drain.rs, and the claim order in
Db::claim_next_due_op in crates/pdfs-core/src/db/ops.rs.
Cause. The drain serialises the ops of one node, but not ops on different nodes that share a
name. Several workers drain at once, so the create of the new b.txt can reach Drive before the
rename that frees the name has landed. Drive then refuses it as taken. drain_local_node waits
when the name is held by a node whose trash is queued (has_pending_trash_named), but not when a
queued rename is about to move it away. It treats the name as someone else's file and lands the
new one under a conflict name.
Fix. With B138's: the create also waits while a queued rename moves a node away from its
name (Core::name_is_held).
Test: the simulation runs fail on it. They counted it instead until the run on 812eb3d
passed every profile on twelve seeds without the count.
B125 — A folder made offline fails with EIO once its listing is invalidated
Status: Fixed.
Found: 2026-10-03, by the simulation runs (sim::run, profile one-client-flaky, seeds 3
and 4). Right after a restart, rmdir y (seed 3), and a write to a file in the client's folder
(seed 4), failed with EIO. The daemon logged "enumerate folder children failed" with Drive's
404 for a local~ folder uid. Each folder had been made while its create was being queued:
offline, or, as for the client's own folder, through B118's fallback when the read-back missed.
Where: Core::ensure_children in crates/pdfs-fuse/src/lib.rs, and everything that calls
State::invalidate_listing.
Cause. A folder whose mkdir is still queued exists only here, so serve_mkdir records it as
listed and empty, and ensure_children never asks Drive about it. State::invalidate_listing
clears that flag for any folder, though, and so do its callers: relocate after a rename, and
relist_offline_listings when the link comes back. The next access asks Drive to list a local~
uid. Drive answers 404, and every lookup in the folder fails with EIO until its mkdir lands.
Right after it lands, Drive may not list the new folder yet (the lag of B118). Then the listing
fails with a 404 for the real uid too.
Fix. A folder still known by its stand-in is listed from the database and never asked of Drive. Until its mkdir lands the folder is only here; once it has, the database holds what Drive does. The second case, a listing by the real uid that Drive does not show yet, needs a create that landed without being read back: the read-back waits up to 1.25 s for Drive to list the folder, longer than the lag of any simulation profile. The runs keep counting both until they show neither again.
Test: a_folder_whose_mkdir_is_queued_lists_after_a_refresh in
crates/pdfs-fuse/src/sim/daemon.rs: with sync paused, a folder is made with a file in it, a
refresh drops its listing, and it is listed again. With PDFS_SIM_KNOWN=fail the flaky profile
fails on it in the simulation runs; seeds 3 and 4 no longer reach it since phase 3.
B124 — After a replacing rename fails, the file it was replacing is gone from the mount
Status: Fixed (unverified).
Found: 2026-10-03, by the simulation runs (sim::run, profile one-client-lan, seed 1). A
rename x/f.txt -> z/d over an existing z/d failed with EIO (B122). Then create z/d
failed with ENOENT, although z/d was back on Drive.
Where: Core::remove_replaced and Core::restore_replaced in crates/pdfs-fuse/src/lib.rs.
Cause. Before the rename, remove_replaced trashes the destination on Drive, drops it from
every mount's state and adds it to hidden, which filters listings. When the rename then fails,
restore_replaced restores the node on Drive but does not take it out of hidden or put it back
in the state, as the ordinary restore does. Until the daemon restarts, the file is on Drive but
missing from the mount, and a create by its name fails.
Fix. restore_replaced takes the node out of hidden, and drops the listing of the folder
it is restored to, and the kernel's entry for its name, in every mount that shows the folder. The
next lookup lists the folder again and finds the node.
Test: the unit test a_replaced_node_put_back_shows_in_the_mount_again. Since B122 is fixed,
the simulation runs no longer reach it through a rename across folders.
B123 — A second write open on a file just written fails with EIO
Status: Fixed (unverified).
Found: 2026-10-03, by the simulation runs (sim::run, profile one-client-lan, seeds 5 and
9). A write at an offset (seed 5) or a create over an existing file (seed 9) failed with EIO.
The daemon logged "refusing write over incomplete queued revision". The one-client-echoes profile,
whose events come late, twice and out of order, found it in four more seeds.
Where: the write open in crates/pdfs-fuse/src/filesystem.rs, the complete flag that
queue_revision in crates/pdfs-fuse/src/lib.rs stamps on a queued revision, and everything that
leaves the content cache without the bytes of the file's base.
Cause. A write open over a queued revision that is not complete is refused by design, because a second partial edit on top of one with gaps cannot be represented. Ordinary cases left a revision incomplete, though:
- A file closed empty never counted as complete.
queue_revisioncompared the authored ranges with[(0, len)], which for a length of 0 is[(0, 0)], while the authored ranges are empty. The truncate path already treated a size of 0 as complete. - The gaps of a partial edit are filled from the content cache, and the cache could be empty or
out of reach then:
- The echo of our own change evicted it, taken for someone else's.
rename_remote,move_remoteand the drain'smove_renamingdid not record an echo, and neither did a create made online, so the create's echo took the one the upload recorded. The feed can also deliver an event twice, or replay one after a batch out of order, and the second copy found the echo already taken. - Drive stamps a link's
ModifyTimeon a rename or a move too. The listing after it moved the file's mtime, and the cache, keyed by mtime and size, no longer matched. - When an upload landed, the drain let go of the staged blob, read the node back, rebased open write handles onto it, and only then cached the blob. A handle released in between found neither the blob nor the cached bytes.
- The echo of our own change evicted it, taken for someone else's.
Either way, the next write open on the file within the drain debounce (DRAIN_REVISION_DEBOUNCE,
2 s) failed, until the revision had landed.
Fix. StagedWrite::covers counts an empty file as complete, at every site that sets the flag.
Core::changed_remote records the echo of every rename and move, and fetch_minted_node the echo
of every create made online. The event sync skips a node update whose id it applied lately
(RecentEvents). State::intern keeps a file's mtime, and the size a listing left out, while its
active revision id stays the same. refresh_after_upload caches the uploaded bytes before it
rebases open handles, and lets go of the staged blob only after.
Test: the unit tests an_empty_write_covers_the_file_and_a_gap_does_not,
a_node_update_seen_again_is_skipped_but_other_events_are_not,
a_renamed_file_keeps_its_mtime_until_its_content_changes and
a_rename_or_move_drops_the_sdk_copy_and_expects_its_echo, and the FUSE tests
a_file_closed_empty_takes_a_second_write_at_once and
the_echo_of_our_own_rename_keeps_the_cached_bytes in sim::daemon. The simulation runs no
longer tolerate it.
B122 — A rename across folders fails with EIO when the source folder holds the new name
Status: Fixed (unverified).
Found: 2026-10-03, by the simulation runs (sim::run, profile one-client-lan, seeds 1, 7
and 10). A rename(2) like x/f.txt -> z/d failed with EIO while x/d existed. The daemon
logged "rename failed" with Drive's AlreadyExists.
Where: serve_rename in crates/pdfs-fuse/src/filesystem.rs.
Cause. Online, a rename that changes both the folder and the name is done in two calls. It
renames in the source folder first, then moves (rename_remote, then move_remote). A sibling in
the source folder with the new name makes Drive refuse the first call, although the destination
is free. B46's follow-up already notes that this path was not converted to the single
move-multiple request the drain uses. If the rename was replacing a file, B124 follows.
Fix. Core::move_rename_remote moves and renames in one move-multiple request, so the name
only has to be free in the destination. serve_rename uses it for every rename across folders,
and the drain uses it in place of its own move_renaming.
Test: the FUSE test a_rename_across_folders_ignores_the_new_name_in_the_old_folder in
sim::daemon. The simulation runs no longer tolerate it.
B121 — A move right after a rename fails with EIO, and a rename across folders takes seconds
Status: Fixed (unverified).
Found: 2026-10-03, in the journal of the account run on flaky Wi-Fi. Each change answered
correctly, but slowly. One rename and move through the mount collected 13 operation failed … method=PUT status=422 warnings over 8.2 seconds, and another collected 7 over 3.9 seconds. Each
one was a request that Drive refused as out of date and the daemon tried again half a second
later.
Where: every remote rename and move: serve_rename in crates/pdfs-fuse/src/filesystem.rs,
the control rename and move_to in crates/pdfs-fuse/src/lib.rs, move_between in
crates/pdfs-fuse/src/relocate.rs, and the drain's renames and move_renaming in
crates/pdfs-fuse/src/drain.rs.
Cause. A rename or move sends the node's current name hash as OriginalHash. The SDK takes
it from its entity cache before the hash in the link it has just fetched. Its own rename and move
do not update that cache entry. Only the event feed's echo of the change drops it, and the daemon
polls the feed every 10 seconds (POLL_INTERVAL). Until then, each rename or move of the same
node sent the old hash. A rename across folders, which is a rename and then a move, was refused
until the next poll, and the retries from B111 hid that as latency. A plain move was not retried
at all. When it came within the poll interval, mv a b && mv b dir/ through the mount failed with
EIO, and pdfs rename followed by pdfs move failed with Drive's "out of date" error.
Fix. Core::forget_sdk_node passes the SDK the event its own change will echo, which drops
the cached node. Every rename, and every move but the drain's, goes through Core::rename_remote
or Core::move_remote. Both drop the entry once the change has landed, and before each retry of a
request Drive called out of date. The drain's move_renaming drops it once its move lands.
Test: the unit tests a_rename_or_move_drops_the_sdk_copy_and_expects_its_echo and
every_remote_rename_and_move_rides_out_a_stale_hash. On the account run, "regression B111" and
"regression B113" should show no status=422 warnings.
B120 — A file queued in a folder that has just landed is put off by five seconds
Status: Fixed (unverified).
Found: 2026-10-02, in the journal of the account run on flaky Wi-Fi. A folder made while the
link was down was queued as a mkdir. When it landed, the journal logged pending create landed … kind=mkdir, and right after it re-interned an authority missing from the local tree; retrying
and pending operation deferred … reason="the node was missing locally and has been re-read" for
a file queued inside it. The file landed on the next recheck. No bytes were lost.
Where: Db::finish_create and Db::remap_local_uid in crates/pdfs-core/src/db/ops.rs, and
adopt_real_uid in crates/pdfs-fuse/src/drain.rs.
Cause. When a queued create landed, finish_create repointed the queued children at the real
uid and deleted the placeholder's row in one transaction. The real node's row was only written
later, by adopt_real_uid, after it had read the node back from Drive. A drain worker that picked
up a child in between found no row for its parent. It read the folder again from Drive and put
the child off by one recheck (DRAIN_ACCESS_RECHECK, 5 seconds).
Fix. adopt_placeholder_row_tx in crates/pdfs-core/src/db/nodes.rs hands the placeholder's
row to the real uid in the same transaction, along with the rows below it, and drops the old row
from the search index. adopt_real_uid writes the server's copy over it later. A row that the
real uid already has, from the event stream, is newer and is kept.
Test: the db tests a_landed_folder_keeps_a_row_until_the_server_copy_replaces_it and
a_landed_placeholder_never_overwrites_the_server_copy. The suite cannot queue a mkdir on demand,
because a mkdir online is made at once even while sync is paused. Instead, --journal-check
now counts re-interned an authority missing from the local tree as an error, so a run that goes
offline catches this race, and B117, if either comes back.
B119 — Reading a file just written through the mount needs the network
Status: Fixed (unverified since the second fix below; "regression B119" passed on one account
run of 2026-10-03 and failed on the next).
Found: 2026-10-02, by the account run on flaky Wi-Fi, which crashed at its very end. The
run's last step reads back a file it wrote to compare digests. The Wi-Fi had dropped, the read
waited for open_revision for 120 seconds and then failed with EIO, and the suite stopped with
an OSError traceback.
Where: the landed-upload path of the drain in crates/pdfs-fuse/src/drain.rs.
Cause. Once an upload landed, the drain kept the staged bytes as cached content only for a pinned file. For any other file it discarded them, and then dropped the reader too. The next read of a file that had just been written therefore downloaded it again, and without a link it could not.
Fix. After the upload, refresh_after_upload reads the node back and now returns it. The
staged blob becomes the file's cached content, linked rather than copied, and stamped with the
mtime and size of the revision just read back, so the cache treats it as current. A file larger
than a sixteenth of the cache budget (LANDED_KEEP_SHARE, 320 MiB of the default 5 GiB) is still
discarded unless it is pinned, so one large upload cannot push the rest of the cache out. Nothing
is kept when the read-back fails or reports a different size. The suite records a failed end-of-run
read-back as a failed check instead of crashing.
Found again on 2026-10-03, by "regression B119" on an account run over a slow link (uploads of a few bytes took 1.4 seconds). The case failed with "block.bin not cached after the upload" at the same second the journal logged that upload as landed.
Cause, second part. The drain retired the op first, and only then read the node back and linked the blob into the cache. For that one round trip the status reported an empty queue for a file that was not yet cached. Anything that waits for an empty queue before reading could still need the network.
Fix, second part. Core::landing_uploads counts an upload from just before its op is retired
until its bytes are cached, and the status adds it to pending_uploads. The status reads the
queue before the count, so it always sees the upload in one of them. The op still goes first: an
op restored after a crash would re-send a revision this daemon has not yet recorded as its own,
and revision_conflict would keep it as a conflict copy.
Test: the drain tests a_landed_upload_stays_cached_unless_it_would_crowd_the_cache_out and
a_landing_upload_counts_until_its_guard_goes, a_landed_upload_stays_pending_until_its_bytes_are_cached
in lib.rs, and the acceptance case "regression B119", which writes two files, waits for the
queue and expects pdfs ls to list both as cached.
B118 — Creating a file or folder through the mount can fail with ENOENT although it was made
Status: Fixed (unverified).
Found: 2026-10-02, by the account run on flaky Wi-Fi. In the first on-demand folder, case 14
failed in its sqlite step with "unable to open database file", and case 15 failed in mkdir
with FileNotFoundError. In both, the file or folder had been made on Drive.
Where: serve_create and serve_mkdir in crates/pdfs-fuse/src/filesystem.rs.
Cause. After a create or mkdir succeeds, the mount reads the new node back to answer the
kernel. Drive can answer the create before its listing returns the new node. The read-back then
came back empty, which fetch_node reports as ENOENT, and the mount passed that on with nothing
in the log. The program saw its create fail for something that existed.
Fix. Core::fetch_minted_node asks again after 250 ms and after one more second
(MINTED_READ_BACK_DELAYS). If the node still cannot be read, the mount queues the create as it
does offline (queue_local_node), and the drain adopts the existing node by its name. The program
gets its file or folder either way. pdfs mkdir uses the same read-back, and when it still misses
it says that the folder was made but is not listed yet, instead of reporting that it does not
exist.
Test: none that replays it on demand, since the gap depends on Drive. The affected acceptance cases ("application workloads" and "block boundaries") fail when it comes back.
B117 — An upload queued for a file pdfs move moved is listed by its raw id and put off once
Status: Fixed (verified 2026-10-03: "regression B117" passed in My files on the account run,
and its journal had no deferral).
Found: 2026-10-02, while verifying B113 live. Three files were moved with pdfs move while
their uploads were queued. Until each upload landed, pdfs sync queue listed it by its node id
instead of its new path. After the resume, the first attempt of each logged pending operation deferred … reason="the node was missing locally and has been re-read", and the upload landed on
the next recheck, about 8 seconds later. No bytes were lost.
Where: settle_source in crates/pdfs-fuse/src/relocate.rs and Core::rename in
crates/pdfs-fuse/src/lib.rs, which forget the node. pending_op_infos in
crates/pdfs-fuse/src/queue.rs, and resolve_unknown_authority in crates/pdfs-fuse/src/drain.rs.
Cause. State::forget deletes the node's row as well as its entry in memory, so the mounts
re-listed it under its new path. Until a listing brought the row back, nothing local knew the
node. pending_op_infos then had no path for the queued op and fell back to its name, or to the
raw uid. The drain's access check found no node either. It re-read the node from Drive, stored it
and put the op off to the next recheck instead of going on with it. pdfs rename forgot the node
the same way.
Fix. pdfs move and pdfs rename move the node instead of forgetting it
(Core::relocate_everywhere). Each mount that has the destination folder moves the node in place
with State::relocate, which keeps its inode and writes its row through. A mount without the
destination drops the node from memory only and keeps the row. When no mount had the node, its
row is moved directly. Both names are invalidated in the kernel, as B114 needs.
Test: the fuse test a_control_move_keeps_the_node_and_its_row_and_drops_the_old_name, and
the acceptance case "regression B117". It pauses sync, rewrites an uploaded file and moves it with
pdfs move. It checks the path pdfs sync queue shows for the upload, and looks in the journal
for a deferral of the node after the resume.
B116 — Trashing a folder keeps the queued creates in its subfolders
Status: Fixed (verified 2026-10-03: "regression B116" passed in My files on the account run).
Found: 2026-10-02, by the first live run of the acceptance case "regression B113". That
version wrote *.part files, which park their create until a rename instead of minting the file
(B70). It moved one of them, still named moved.part, into the subfolder b113/moved here and
then hung, waiting for a queue that held a parked op. The run was interrupted, and its cleanup
trashed the test root with pdfs rm and deleted it permanently. The parked create stayed queued
and kept its staged bytes. The cleanup reported it as left over (queued create #2 for …/b113/moved here/moved.part).
Where: drop_doomed_ops in crates/pdfs-core/src/db/ops.rs, which
Core::discard_queued_ops runs when a folder is trashed.
Cause. The ops removed with a folder are those for the folder and for the nodes queued below
it through pending_op.parent_uid, plus the revisions of every file the node tree places below
it (B102). A create for a file inside a real subfolder is neither: its parent is not a queued
node, and the tree only contributes revisions. Such a create comes from a transient name, or
from a file created offline in a folder that already exists on Drive. A parked one stays queued
for good, holding its bytes. One that is due would run into a trashed parent, and by the code the
drain then re-homes it to the root (parent folder is gone remotely; creating in the root instead), so the file would reappear in My files after its folder was trashed. Only the parked
case has been seen.
Fix. drop_doomed_ops also takes the queued creates and mkdirs whose parent is anywhere in
the node tree below the folder, together with everything queued below those.
Test: the db test trashing_a_folder_drops_the_queued_creates_in_its_subfolders, and the
acceptance case "regression B116", which replays the due case. With sync paused, it writes a
*.part file two folders below a new folder and renames it to its final name, so its create is
queued and due. It then trashes the folder with pdfs rm and expects the create gone. It restores
the folder before sync resumes, so a create that wrongly stayed lands where the file was written,
and fails if the file shows up in the root anyway.
B115 — pdfs ls shows a file's encrypted size
Status: Fixed (verified 2026-10-03: "regression B115" passed in My files on the account run).
Found: 2026-10-02, while verifying B88 live. pdfs ls ~/Downloads/pdfs-b88-test listed the
6-byte note.txt as 57 bytes, and listed tickets.pdf as 202316 bytes where the mount's stat
said 202257. The acceptance case "open B115" then showed it for every file it wrote, in My files
and in an on-demand folder, right after the upload and again after pdfs refresh: 6 bytes listed
as 57, 4097 as 4150, and 4 MiB + 7 bytes (4194311) as 4194421. The mount's stat had the real
sizes throughout.
Where: node_size in crates/pdfs-fuse/src/lib.rs and crates/pdfs-core/src/db/nodes.rs.
Cause. node_size reports the node's claimed_size and falls back to
total_size_on_storage when it has none. That is the size on Drive's storage: encrypted blocks
plus their signatures. A folder is first listed without claimed sizes, to be quick, and the real
sizes are fetched for the whole listing in the background. stat through the mount waits for that
fetch. pdfs ls and the app's file browser read the listing straight away, so they showed the
encrypted sizes until it landed.
Fix. Core::list_dir, which serves both, waits for the listing's size fetch first
(settle_listing_sizes), for at most SizeUpgrade::WAIT (10 seconds), as stat does. Offline it
does not wait, since the fetch cannot succeed.
Test: the fuse tests a_listing_for_the_cli_settles_sizes_before_reading_them and
a_listing_waits_for_the_whole_size_batch_but_not_forever, and the acceptance case
"regression B115". It compares the sizes pdfs ls reports with the written ones, after the upload
and again after pdfs refresh.
B114 — A name pdfs rm or pdfs rename took away still resolves for up to 30 seconds
Status: Fixed (verified 2026-10-03: "regression B114" passed in My files on the account run).
Found: 2026-10-02, while verifying B88 live. After pdfs rm ~/Downloads/pdfs-b88-test the
folder was gone from ls ~/Downloads, but ls -d ~/Downloads/pdfs-b88-test still found it for
up to 30 seconds. ~/ProtonDrive behaves the same, so B88's routing did not cause it.
Where: Core::rename and Core::delete in crates/pdfs-fuse/src/lib.rs, and the TTL the
kernel gets for entries in the same file.
Cause. Both change Drive, forget the node and drop the parent's listing, but never tell the
kernel. The kernel cached the name's lookup for TTL, 30 seconds, and answers from that cache
until it expires. A rename or unlink through the mount has no such gap, because the kernel made
that change itself. pdfs move already sends inval_entry for its destination
(settle_dest in crates/pdfs-fuse/src/relocate.rs); its source was not checked.
The account run on flaky Wi-Fi on 2026-10-02 hit it in another place. "regression B111" renamed
a file with pdfs rename and straight back, and then failed with FileNotFoundError: the
kernel still pointed the first name at the inode the rename had forgotten.
Fix. pdfs rm forgets the node in every mount and sends each kernel session a delete notice
for the name (Core::forget_everywhere). pdfs rename and pdfs move invalidate the old and the
new name (Core::relocate_everywhere, see B117). The notices go out after the mount's lock is
released.
Test: the acceptance case "regression B114" removes one file with pdfs rm and renames another
with pdfs rename, then checks at once that neither old name resolves. It reproduced in My files
and in an on-demand folder before the fix.
B113 — A new file moved before its upload drains lands as a conflict copy of itself
Status: Fixed (verified 2026-10-02).
Found: 2026-10-02, on the production account. aria2 downloaded two seasons into the on-demand
~/Downloads. The episodes were moved into Videos/anime/… while their uploads were still
queued. 45 of 48 landed as (sync-conflict …) copies next to a 0-byte original, each with
the remote revision changed under the queued write (expected 0 bytes at mtime 1790955809, found 0 at 1790956593). The three that drained before the move uploaded normally.
Where: serve_create in crates/pdfs-fuse/src/filesystem.rs, adopt_real_uid in
crates/pdfs-fuse/src/drain.rs.
Cause. The handle a create opens is based on the empty file the server mints, but it was
given no revision id. B69 keys the drain's conflict check on the revision id, and without one it
falls back to (mtime, size). The node's mtime is the link's modify time, which the server
advances on a move or rename. So a new file moved before its first write drained looked like
another device had changed it. A file created offline had the same gap in another place: a handle
still open when its queued create landed kept the placeholder's local-clock mtime and no revision
id.
Fix. WriteHandle::created takes the minted node's revision id along with its mtime. When a
queued create lands, adopt_real_uid rebases the handles still open on it onto the new node,
as refresh_after_upload already does after a sealed upload. Both go through
WriteHandle::rebase_onto.
Verified live (2026-10-02) on the installed daemon, in a scratch folder in ~/Downloads. A
pdfs move advanced an empty file's server mtime from 1790962386 to 1790962408, the change the old
check took for another device's edit. With sync paused, three 16 MiB files were written and moved
into a subfolder while their uploads were queued. After the resume all three logged pending upload landed, with no conflict warning or copy, and they read back at their exact sizes. Each upload was
put off once after the move (the node is missing locally and was re-read) and landed on the
retry about 8 seconds later. Meanwhile pdfs sync queue showed raw ids instead of the new paths.
Both are filed as B117. Not verified live: a file created offline that is still open when its
create lands. Only the unit test covers that half. The acceptance case "regression B113" replays
the online order through the mount, with a rename and a move, and passed in My files and in an
on-demand folder.
B112 — Folders shared by the same person disappear from Shared with me
Status: Fixed (unverified).
Found: 2026-09-30, reported in #25. An account with several accepted shares from one other
account listed only some of them, in both the control listing and the FUSE mount. The journal
showed skipping undecryptable node ... failed to decrypt: missing key. Each missing folder
opened on the Proton Drive website and resolved when the SDK was asked for it alone.
Where: proton-drive-rs 0.7.1, enumerate_nodes_detail, which serves both
enumerate_nodes and enumerate_nodes_light. The app hit it in list_shared_with_me
(crates/pdfs-fuse/src/sharing.rs) and in the shared-root listing in
crates/pdfs-fuse/src/lib.rs.
Cause. The SDK cached parent keys per batch by the link's parent id. Every share root has
no parent, so all roots on one volume shared the None entry, and the roots after the first were
decrypted with the first root's share key (proton-sdk-rs#2).
Fix. Require proton-drive-rs 0.7.2, which names each key Parent(folder) or Root(link).
No app code changes. The app-side workaround in #25, one SDK call per root, is not needed.
Unverified: no account with several shares from one owner has listed through 0.7.2 yet.
B111 — Renaming a file back right after renaming it fails with EIO
Status: Fixed (unverified).
Found: 2026-09-29, by the acceptance case "unusual but legal names", on My files and on the
on-demand folder. The case renames each file to renamed-N and back at once. One rename back per
run failed with EIO, and the journal showed rename failed ... proton api error InvalidRequirements (http 422): This file or folder was out of date, rename failed.
Where: every remote rename: serve_rename in crates/pdfs-fuse/src/filesystem.rs,
drain_rename_in_place in crates/pdfs-fuse/src/drain.rs, and the control rename in
crates/pdfs-fuse/src/lib.rs.
Cause. The SDK's rename_node sends the node's current name hash as OriginalHash. Right
after a rename of the same node, it still sent the old one. Drive then refuses the request as out
of date. The daemon reported this at once as EIO, although the same request succeeds a moment
later. The old hash comes from the SDK's entity cache, not from the server (see B121).
Fix. All remote renames go through Core::rename_remote. It retries a rename that fails with
InvalidRequirements up to 20 times, half a second apart, the same window the move after a
rename already uses. Any other error still returns at once.
Test: the acceptance case "regression B111" renames a file there and back eight times through
the mount, and three times with pdfs rename, and checks its name and content after each round.
B110 — A read that settles a provisional size deadlocks on its own page lock
Status: Fixed (unverified).
Found: 2026-09-29, by the acceptance case "application workloads", which timed out on My
files and on the on-demand folder on a fast network. The journal showed a fuse worker has held the same job for a long time worker=pdfs-fuse-8 job=read growing past 700 s with no network
error. /proc/<pid>/task/*/wchan put that worker, and one more, in __folio_lock, while the
daemon had no file on any mount open.
Where: crates/pdfs-fuse/src/lib.rs (Core::apply_size_upgrade).
Cause. B100 made read settle a provisional size before it fetches, through
resolve_provisional_size and apply_size_upgrade. That function tells the kernel about the new
size with inval_inode(ino, 0, 0), which also drops the file's page cache. The kernel holds the
page lock of the read being served until the daemon replies. The invalidation waits for the same
lock, on the thread that has to send the reply, so neither ever finishes. The process that
issued the read hangs, and so does anything else that touches those pages.
Fix. The size upgrade invalidates attributes only (offset -1). A size upgrade does not
change content, so the page cache has nothing stale to drop. A daemon already stuck this way
needs a restart.
B109 — A re-listing shrinks a file that is open for writing, and SQLite reads it as malformed
Status: Fixed (unverified).
Found: 2026-09-29, by the acceptance case "application workloads" on My files. It failed
several runs in a row with database disk image is malformed at the SELECT COUNT(*) right
after a 500-row insert, and the daemon logged nothing. A diagnosis added to the case showed
workload.db at 0 bytes while SQLite still held it open with 16 KiB written.
Where: crates/pdfs-fuse/src/state.rs (State::keep_open_write_size, called from
intern_mem_with_access), the local-uid remap in crates/pdfs-fuse/src/drain.rs, and the size
upgrade in crates/pdfs-fuse/src/lib.rs.
Cause. An online create seals an empty file on Drive. Until the last close, the written
bytes live only in the open handle, and the entry's size is the handle's. Any re-listing of the
parent re-interned the node from Drive, which replaced that size with 0. Such re-listings come
from event echoes and sibling creates, and SQLite's journal is one. The kernel then got size 0
from getattr and ended SQLite's reads at EOF without asking the daemon, so the database read
back as empty pages. Queued writes were already protected (B11); open handles were not.
Fix. A node that has an open write handle keeps the handle's size and its local mtime when a listing, a size upgrade or a create landing replaces its entry. After the last close, the queued write takes over as before.
B108 — A file deleted during a mirror pass puts the sync folder in error
Status: Fixed (unverified).
Found: 2026-09-29, by the account run of scripts/fuse-acceptance.sh --quick. It stopped with
sync folder … entered error before the managed-move phase. At the same time the journal
showed sync: step failed; continuing error=open …/workloads/tree.tar: No such file or directory
for eleven files. The contract suite had removed its workloads tree while the mirror pass
was still uploading it.
Where: crates/pdfs-fuse/src/sync.rs (upload_new, upload_revision, open_to_upload).
Cause. A pass plans its uploads from a scan and runs them later. When a planned file was
deleted in between, opening it failed with NotFound. That counted as a failed step, so the pass
ended in error and logged "item(s) failed; will retry" for files nobody wanted any more.
Fix. An upload whose local file is gone when it starts is skipped (Applied::Vanished) and
does not count as an error. The next pass sees the deletion and handles it like any other.
Other open failures are still errors.
B107 — One read on a stalled connection hangs the reading program for half an hour
Status: Fixed (unverified).
Found: 2026-09-29, by the acceptance case "application workloads" on My files, which hit its
180 s timeout. The journal showed a fuse worker has held the same job for a long time worker=pdfs-fuse-10 job=read from age_secs=124 up to age_secs=2059. Event polls in the
same window failed with HTTP transport error after 91-103 s.
Where: crates/pdfs-fuse/src/reads.rs (Core::read_range_remote, READ_FETCH_TIMEOUT).
Cause. The SDK bounds each request (30 s for the API, 300 s for a storage block) and retries
failed ones, but nothing bounds a whole read. A read can need a block table, several blocks and a
repair, each with its own retries. On a connection that stalled without dropping, one read
held its FUSE worker for about 34 minutes, and the program that issued it could not be
interrupted for that time.
Fix. The network part of a read now has a 120 s deadline. When it passes, the read fails
with EIO, the file's revision reader is dropped so the next read starts fresh, and blocks that
already arrived stay cached for the retry. A single read is at most a few 4 MiB blocks, so this
only fails a read on a link slower than about 70 KB/s. A block repair that has to download a
large file whole can also hit the deadline on a slow link; pinning the file avoids that.
The first version of this fix built the timer on the FUSE worker thread, outside the runtime.
Every read that went to the network panicked with there is no reactor running, and the kernel
got EIO. The next account run failed 11 cases on My files this way. The timer is now built
inside block_on, and a unit test runs that pattern from a thread without a runtime.
B106 — Saving over a file that was just read fails with "already exists"
Status: Fixed (unverified).
Found: 2026-09-29, by the acceptance case "application workloads" on an on-demand folder,
after installing the B104 fix: OSError: [Errno 5] from os.replace in the editor-save loop,
at the third save. The journal had rename failed … AlreadyExists (http 422): A file or folder with name already exists from rename_node.
Where: crates/pdfs-fuse/src/lib.rs (Core::remove_replaced,
Core::withdraw_queued_ops), crates/pdfs-fuse/src/filesystem.rs (trash_child,
finish_release), crates/pdfs-fuse/src/drain.rs (Core::drain_revision).
Cause. A regression from B104. The save loop writes a temporary file, renames it over the
document, and reads the document back. close() returns before the daemon sees the release, so
at the next save the replaced file still counts as open. B104 then left its queued upload in
the queue until the final close, and the drain could send it after the file had been trashed.
The next rename onto the name was refused. unlink of an open file had kept its queued upload
the same way.
Fix. A file trashed or replaced while open now loses its queued ops at once, and an upload already under way is cancelled. Only the staged bytes stay, because the open handle reads them, and they are deleted at the final close. A drain worker that claimed the op before it was withdrawn checks that the op still exists before it sends any bytes.
B105 — A file deleted while its upload is in flight comes back as a conflict copy
Status: Fixed (unverified).
Found: 2026-09-29, in the daemon journal of an acceptance run on the account. During the
SQLite workload two queued write conflicts; keeping a conflict copy warnings named files called
recovered-XrPUSW5W and recovered-JbZM_pWK, with the reason the file was trashed remotely.
Each was followed by pending upload failed; will retry and No such file or directory.
Where: crates/pdfs-fuse/src/drain.rs (Core::drain_revision),
crates/pdfs-core/src/db/ops.rs (Db::op_exists).
Cause. SQLite writes, syncs and deletes its journal file within milliseconds. The sync
queues an upload of the journal. The delete trashes the node and drops that queued upload, but a
drain worker that already holds the op carries on. It then asks the server about the node,
finds it trashed, and treats the trash as another device's change: it tries to upload the bytes
as a conflict copy. The tree has already forgotten the node, so the copy is named recovered-…
and placed in the Drive root. Here the blob was already gone and the attempt failed, but with
the blob still on disk the copy would land.
Fix. Before keeping a conflict copy, the drain checks that its op is still queued. When the op is gone, the write was withdrawn by the delete, or superseded, while the worker held it, and it is dropped. The check goes by op id, which the next op queued can take over (B133).
B104 — Reading a file that was replaced by a rename fails with EIO
Status: Fixed (unverified).
Found: 2026-09-29, by the acceptance case "rename patterns" on My files and on an on-demand
folder: OSError: [Errno 5] Input/output error at
os.pread(reader, 64, 0) after os.replace(replacement, target) with reader open on target.
The mirror folder, a local disk, passed.
Where: crates/pdfs-fuse/src/lib.rs (Core::remove_replaced).
Cause. A rename over an existing file first trashes that file, because Drive has no atomic
replace. That step forgot the replaced inode outright and evicted its cache and reader, even
when a handle was still open on it. The next read through that handle found no inode and
failed. rename(2), like unlink(2), keeps a replaced file readable through the handles
already open on it. The unlink path (trash_child) already did this, but the rename path
did not.
Fix. remove_replaced now handles an open victim as unlink does. It keeps the inode as
an unlinked entry, and leaves its cache and reader to the final close. That applies both to a
node the server knows and to one whose create is still queued. Its queued ops at first also
waited for the final close; B106 moved them back to the trash.
B103 — One stuck trash refresh freezes the trash listing until the daemon restarts
Status: Fixed (unverified). The hang itself has not been reproduced on purpose.
Found: 2026-09-29, while timing the acceptance suite's wide-directory case. Its cleanup trashed
the test folder, then waited for it to show up in the trash so it could delete it permanently. It
never did. From 06:55 UTC on, every trash listing logged trash refresh still running; answering with what has materialized and returned the same 362 entries. No trash refresh: enumerated line
followed, so the first server call of the refresh never returned.
Where: crates/pdfs-fuse/src/lib.rs (Core::refresh_trash, within_trash_call).
Cause. Trash refreshes are single-flight: a request joins the refresh already running instead of starting another. The refresh awaited its server calls with no deadline, so one call that never returned kept the flag set forever. Every later request joined it, waited out its 20 s budget, and answered with the stale listing.
Fix. Each server call of a refresh (enumerate_trash_node_uids, and each enumerate_nodes
chunk) fails after TRASH_CALL_TIMEOUT (120 s). The refresh then ends with an error, clears the
flag, and the next request starts a fresh one.
B102 — Removing a folder leaves the queued uploads of the files inside it
Status: Fixed (unverified). No daemon carrying the change has run the acceptance suite yet.
Found: 2026-09-29, after the account run that found B101. Its cleanup removed both test roots
and reported "nothing left behind", but pdfs sync queue still listed two revision ops for
pdfs-acceptance-…/boundaries/reshaped, each holding about 8 MiB of staged bytes. They retried
against the trash forever ("the share has not been writable for 300s").
Where: crates/pdfs-core/src/db/ops.rs (drop_doomed_ops, used by
Db::replace_ops_with_trash and Db::delete_ops_for_uid), crates/pdfs-fuse/src/lib.rs
(Core::delete, Core::release_dropped_ops).
Cause. Two gaps.
- The queue could not see which files a folder holds. Removing a node dropped the ops for it
and for anything queued beneath it, found through
pending_op.parent_uid. Onlycreateandmkdirops carry a parent, so arevisionof a file that already existed remotely was never found, however deep in the removed folder it sat. pdfs rmdropped nothing at all. The control socket's delete trashed the node remotely but left every op queued for it, unlike an unlink through the mount.
Fix. Removal also follows the node tree (nodes.parent_uid) and drops the revision ops of
every file below the node, with their blobs, pending entries and in-flight uploads. Queued renames
below it are kept: one that moved a node into the folder still has to land for the node to go to
the trash with it. pdfs rm now drops the queued ops once the remote trash succeeds. The
acceptance janitor lists the daemon queue after cleanup and reports any op still naming one of its
roots as a leftover.
Regression test: trashing_a_folder_drops_the_queued_revisions_of_files_below_it.
The acceptance case "regression B102" pauses sync, rewrites a file two folders down in each of
two folders, and removes one folder with pdfs rm and the other with rm -r through the mount.
Neither revision may stay queued.
B101 — A shrunk file written past its end gets old bytes back, or fails with EIO; a write during a create's upload is lost
Status: Fixed (unverified). No daemon carrying the change has run the acceptance suite yet.
Found: 2026-09-29, by the new acceptance case "block boundaries and overwrites that change
the block count". It failed with Input/output error in My files and in an on-demand folder, every
time. The journal had the revision's block table disagrees with the node size ... fsize=8388611 reader_size=0, then pending upload failed; will retry ... gap-fill from base failed: Errno(5)
on every retry.
Where: crates/pdfs-fuse/src/state.rs (WriteHandle::record_write),
crates/pdfs-fuse/src/drain.rs (Core::drain_local_node, Core::retire_create),
crates/pdfs-fuse/src/lib.rs (Core::enqueue_staged_write),
crates/pdfs-core/src/db/ops.rs (Db::finish_create).
Cause. Two defects.
- The hole of a write past the end was not authored. A write handle tracks which ranges it
wrote; everything else below the base size is filled from the base when it is committed. A
grow through
setattrmarks its new tail as written, but apwritepast the end did not mark the hole it leaves. Afterftruncate(fd, 4 MiB)andpwrite(fd, data, 4 MiB + 10), bytes 4 MiB to 4 MiB + 10 were refilled from the old base instead of reading as zeros. That silently restored truncated bytes. When the base was itself a queued revision the remote did not have yet, the fill read a revision of the wrong size, and B84's check failed it:EIOon read, and a drain that retried forever. - A write that closed while its file's create uploaded was lost. The create of a new file
drains on its own, often while the file is still open. A write that closes meanwhile attaches
its blob to the create's queue row. When the create landed, the drain deleted that row by id,
so the server kept the older bytes (often an empty file) and nothing was queued to replace
them. A handle still open after the create landed carried the placeholder uid, found no create
to attach to, and failed its
closewithEIO.
Fix. A write past the end authors the hole (WriteHandle::record_write).
Db::finish_create retires a landed create in one transaction: if the row now holds a newer blob
than the one uploaded, it becomes a revision of the real uid instead of being deleted. The drain
does this under the pending lock, which enqueue_staged_write holds from attaching a blob until
it records the pending entry, and then repoints open write handles to the real uid. A write that
still finds the create gone waits for its inode's real uid (Core::landed_uid) and queues a
revision against it.
Regression tests: a_write_past_the_end_authors_the_hole_it_leaves,
a_write_attached_while_its_create_uploads_becomes_a_revision,
a_create_whose_blob_landed_is_retired, and the live acceptance cases "block boundaries and
overwrites that change the block count" (now also read back after the queue drains) and
regression B101.
B100 — A freshly copied file reads as EIO or 53 bytes too big
Status: Fixed (unverified). No daemon carrying the change has read a file with a provisional
size yet.
Found: 2026-09-29, checking a 222-file copy into Documents/pi-backup on 2.7.0. find -printf %s showed 10 small files 52 to 59 bytes bigger than their originals. cat of one of them failed
with Input/output error, and the journal had the revision's block table disagrees with the node size ... fsize=221 reader_size=169 followed by repair download did not match the node size. The node's DB row and claimed_size were correct (169). A few minutes later the same files
reported the right size and matched the originals byte for byte.
Where: crates/pdfs-fuse/src/filesystem.rs (read, serve_open),
crates/pdfs-fuse/src/lib.rs (Core::upgrade_sizes, Core::run_size_upgrade).
Cause. Three things together.
- A listing built from the cheap enumeration (B12) has no
claimed_size, so a file reports its ciphertext size until the size upgrade lands. The upgrade's own documentation said reads were unaffected because the revision reader carries its own size. - B84 made that untrue.
read_blocknow treats a revision that disagrees with the size it was given as damage: it repairs from the whole-file download, and when that does not match either, it fails the read. With a provisional size the two never match, so every read failed. getattrnormally settles the size before answering (B14), which kept this rare. But onlyMAX_SIZE_UPGRADES(8) folders may upgrade at once, and past that capupgrade_sizesreturned nothing. Thestatwas answered with the ciphertext size at once, and nothing retried the folder until the nextstat. A recursive walk over many fresh folders goes past the cap.
A write open had the same fault: its handle started at the provisional size, so the file would gain the encryption overhead and reading the base would fail.
Fix. read and a write open settle a provisional size before using it
(Core::resolve_provisional_size, one node fetch, on the worker that already goes to the
network). A folder past the cap is now queued instead of dropped. When a running upgrade ends,
its thread takes over the oldest queued folder, so at most 8 threads still run.
Regression tests: a_folder_past_the_cap_is_queued_not_dropped,
a_finished_batch_hands_its_thread_the_oldest_queued_one. The read and open paths need a live
mount and have no unit test; the acceptance case regression B100 covers them on a real mount.
B99 — Conflict copies in synced folders are logged again on every start and shown as "Resolved"
Status: Fixed (unverified). No daemon carrying the change has been restarted against the
affected copies yet.
Found: 2026-09-27. The Activity page listed the same six conflict copies every day, e.g.
Conflict [EMBER] Oshi no Ko S2 - 12 (sync-conflict 1786815864).mkv with Differs from … · 24 times since Sep 20, 13:13 · Resolved. find showed all six copies still on disk, in ~/Videos and
~/Documents. Both are on-demand sync folders. pdfs conflicts answered No conflicts.. The
journal showed the daemon mounting at 17:29:33, and the entries were stamped 17:30. That is the
sweep's 30-second warmup.
Where: crates/pdfs-fuse/src/sweep.rs (flag_conflict, sweep_conflicts_once),
crates/pdfs-fuse/src/conflicts.rs (list_conflicts), crates/pdfs-fuse/src/control.rs
(ResolveConflict).
Cause. There were two faults.
- The sweep kept its "already flagged" set only in memory. Every daemon start forgot it, so the first pass logged every open conflict again.
- The sweep flags copies anywhere in the account's own tree.
list_conflictsonly reported copies below My Files, and resolving only accepted paths under the primary mount. A copy in a synced folder was logged but never listed, so the GUI read it as resolved. It also could not be resolved from the GUI or the CLI.
Fix. Core::conflict_path names a copy by its mountpoint-relative path under My Files, or
by its absolute local path inside a synced folder. It reuses the roots that search already
resolves hits through. The listing uses it, and the sweep flags only copies it can name. So the
feed never holds a conflict the listing cannot answer for. ResolveConflict routes an absolute
path to the on-demand mount that serves it. For a copy in a mirror folder, it says to edit the
local file instead. Before logging, flag_conflict checks the activity log
(Db::activity_has), so a restart does not log the same copy again. The copy's name carries its
own timestamp, so the name identifies the copy. No schema change.
Regression test: activity_has_finds_only_matching_kind_and_target.
B98 — A superseded upload spins its drain thread forever and shows as "Uploading" at 0 bytes
Status: Fixed (unverified) — the daemon carrying the change has not been driven against a
superseded upload yet.
Found: 2026-09-23. The Sync page showed Unconfirmed 654050.crdownload — 0 bytes of 23.8 MiB
for hours after Brave had finished and renamed the download. pdfs transfers listed the same
entry, pdfs diagnostics showed every worker idle, and /proc/<pid>/task showed pdfs-drain-1
in state R at 100% CPU, holding a deleted staging blob open.
Where: crates/pdfs-fuse/src/transfers.rs (CountingReader::read).
Cause. A write that lands while a revision is on the wire cancels that upload through the
flag CountingReader checks before every read. The reader reported the cancel as
ErrorKind::Interrupted. Every read loop treats that kind as "try again", including the SDK's
read_full_block, and the flag never clears, so the loop never ended. The TransferGuard was
never dropped, so the transfer stayed in the registry, and the drain thread was lost until the
daemon restarted. A browser download is the natural trigger: the .crdownload file keeps growing,
so each new write supersedes the upload of the previous one.
Fix. The cancel is reported as ErrorKind::Other, which the SDK turns into an error.
drain_revision already recognises a cancelled upload and returns without counting a failure.
Regression test: a_cancelled_upload_fails_its_read_instead_of_asking_for_a_retry.
The acceptance case "regression B98" appends to a 64 MiB file while its upload is on the wire. The
superseded transfer has to end, and the file has to land at its full size. The case skips when the
upload finishes before it is seen.
B97 — Photo refreshes stack up on the control socket, and shutdown turns the resolve pass into a burst of "undecryptable photo" warnings
Status: Fixed (unverified) — the daemon carrying the change has not been driven against a real
library yet.
Found: 2026-09-19, while reading the journal for issues that were still open. pdfs diagnostics
had reported a Refresh running for 23 minutes with in_flight at 5, and every mount that ended
in a systemctl --user restart was followed by ~95 skipping undecryptable photo warnings.
Where: crates/pdfs-fuse/src/control.rs (Refresh handler), crates/pdfs-fuse/src/photos.rs
(refresh_timeline), crates/pdfs-fuse/src/diagnostics.rs (request_kind).
Cause. Two unrelated faults with one symptom each.
The stall: every Refresh arm is individually cheap — a directory refresh resolves and
invalidates one listing, a photos refresh clears one state key and spawns the real work — but
nothing stopped several of them running at once. A front end that asks again while the first
refresh is still going gets a second handler that does the whole thing again: invalidate_photos,
optionally photos_mark_all_unresolved over the entire timeline, invalidate_albums. Each of
those takes the database lock the already-running refresh holds for minutes at a time, so the
duplicates queue behind it, hold a control handler slot, and appear as a wedged Refresh. The
report could not even say which view was wedged: request_kind reads only serde's external tag,
so all three scopes logged as a bare Refresh.
The warning burst: refresh_timeline resolves photo metadata in chunks and keeps going through
teardown. The runtime is cancelled under it, every decrypt task in flight fails with "task was
cancelled", and the SDK logs one skipping undecryptable photo per node. Nothing was corrupt —
the library was fine on the next mount, because photos_replace preserves prior metadata and
leaves resolved_at null so the next refresh re-reads exactly those photos.
What changed. A refresh scope is now single-flight: RefreshGuard holds the scope key
(dir:<path>, trash, photos, photos:full) for as long as one is being served, and a second
request for that same key is answered refreshing immediately instead of repeating the work. A
non-full photos refresh is additionally skipped while timeline_refreshing() is true, since that
work outlives the request that started it. request_kind now names the scope, so the diagnostics
report says Refresh(Photos) rather than Refresh. And the resolve loop stops asking for chunks
once Shutdown::is_stopping(), logging stopped and the resolved-versus-wanted counts on the way
out.
B96 — Staged bytes that name no node are never retired, and are re-announced on every boot
Status: Fixed (unverified). Found: 2026-09-19, in the journal: the same "staged bytes belong to no queued upload" warning appeared once per blob per mount, forever. The staging directory held 138 blobs / 575 MiB, of which 95 (~445 MiB) referenced no node at all; the oldest were two months old.
Where: crates/pdfs-fuse/src/lib.rs (reconcile_staging).
Cause. Startup reconciliation is deliberately conservative: staged bytes are the only copy of an un-uploaded write, so a blob it cannot match to a queued op is kept rather than deleted. But "kept" had no end. A blob whose node is gone from the database can never be matched by any future boot, so the warning was both permanent and per-blob, and the bytes accumulated with no ceiling.
What changed. A blob that names no node and whose mtime is older than STAGING_ORPHAN_RETAIN
(30 days) is retired, with its size logged. A month is long enough that a write stranded by a bug
can still be recovered by hand. Younger orphans are still kept, but they are now counted and
reported as one aggregated warning with one example path, not one line each.
B95 — A real index.tmp is treated as a browser tempfile and parked forever
Status: Fixed (unverified).
Found: 2026-09-19, while working out why 43 uploads were parked. Most of them were
index.tmp and preview.tmp files from a Google Photos Takeout import — ordinary files whose
final name simply ends in .tmp.
Where: crates/pdfs-core/src/syncignore.rs (TRANSIENT_SUFFIXES).
Cause. One suffix list served two mechanisms with very different consequences. On a mirror
mount, calling a name transient means "do not reconcile it yet", which is reversible and costs
nothing if wrong. On an on-demand mount it means "park the create until the writer renames it"
(B70) — and a file whose real, final name ends in .tmp is never renamed, so the park stands
forever and the bytes never go up.
What changed. .tmp and .temp are out of TRANSIENT_SUFFIXES. The remaining suffixes
(.crdownload, .part, .partial, .download, .swp, .swx) belong to writers that always
rename. A .tmp write from a program that does rename is still handled — by the rename itself,
which supersedes the create.
Test: the acceptance case "regression B95" writes index.tmp and preview.temp and waits for
the queue, which fails at once on a parked op.
B94 — A parked create has no exit but the rename, so an abandoned one is stuck forever
Status: Fixed (unverified).
Found: 2026-09-19, while reading the journal: 43 rows in pending_op carried the
PARK_UNTIL sentinel, some for weeks. pdfs status reported them as parked uploads that never
moved.
Where: crates/pdfs-core/src/db/ops.rs (parked_create_ops, PARK_EXPIRY_MS),
crates/pdfs-fuse/src/drain.rs (sweep_parked_creates).
Cause. B70 parks a transient create by writing PARK_UNTIL into next_attempt_at, and the
only thing that clears it is the transient-to-final rename (Db::set_create_hold). If the rename
never happens — the writer crashed, the user deleted the temp file, or the name was never
transient in the first place (B95) — the row is invisible to claim_next_due_op and stays in the
queue permanently, holding its staged blob.
What changed. The primary drain worker sweeps parked creates when the queue is idle, which is
exactly when a park that has outlived its writer matters. A parked create whose node is gone from
the database has its queued ops discarded. One whose node still exists, that is older than
PARK_EXPIRY_MS (1 hour) and that nobody has open, is un-parked rather than dropped — drain
invariant 1 says staged bytes are only dropped once their op has landed, so the sweep lets the
bytes through under the transient name instead of destroying them. An open node is left parked
whatever its age: it is still being written.
Test: the acceptance case "regression B94" deletes a *.part file whose create is parked and
waits for the create to leave the queue. That is the delete exit. The hour-long expiry is not
driven.
B93 — A photo whose node would not resolve lost its name, and with it its tab and its group
Status: Fixed (unverified) — found and fixed while making the timeline refresh incremental
(B92); no library with an undecryptable photo in it has been driven against the fix.
Found: 2026-09-19, by reading photos_replace before making a refresh skip photos it has
already resolved. Skipping one means writing a TimelineRow that carries only the uid and the
capture time, which is exactly the row a failed resolve already produced.
Where: crates/pdfs-core/src/db/photos.rs (Learned, photos_replace).
Cause. photos_replace keeps what a refresh could not learn — ratio, thumbnail verdict, media
type, favourite flag, content hash, photo relation all survive a None. The name did not:
Learned never held it, the SELECT never read it, and the INSERT bound row.name straight
through. So a photo the SDK could not decrypt (it logs and skips the node) had its stored name
overwritten with NULL on every refresh. That is not a cosmetic loss — both the tab split and the
grouping are derived from the name. PhotoKind::classify is extension-first, so a nameless RAW
file classifies as a still photo and leaves the Raw tab; name_stem returns None, so the RAW and
the JPEG of one shot stop being one shot; and redate skips nameless photos outright.
What changed. Learned carries the name like everything else, and the kind and the group stem
are computed from the carried name rather than the incoming row. A row that was resolved is now
authoritative instead: its fields are taken as they are, so a photo unlinked from its main on
another device finally loses its main_uid — a carry could never express "the server says none".
B92 — Refreshing the gallery reported "Resource temporarily unavailable"
Status: Fixed (unverified) — the daemon carrying the change has not been driven against a real
library yet.
Found: 2026-09-19, while checking why a 12,445-photo timeline still showed the pre-grouping
layout after pressing Refresh. pdfs refresh photos failed after about two minutes with:
Caused by:
0: io: Resource temporarily unavailable (os error 11)
1: Resource temporarily unavailable (os error 11)
Where: crates/pdfs-fuse/src/control.rs (RefreshScope::Photos),
crates/pdfs-core/src/control.rs (READ_TIMEOUT, the new PhotosRefreshStatus),
crates/pdfs-gui/src/app/main.rs (refresh_then), crates/pdfs-cli/src/main.rs (cmd_refresh).
Cause. Refresh { Photos } awaited refresh_timeline() before replying, on the reasoning that
somebody who presses Refresh wants the truth rather than a stale page. But the refresh resolves
every photo's node in batches of TIMELINE_ENRICH_CHUNK (200), so a 12,445-photo library is 63
server round-trips — minutes of work. The client's read bound is READ_TIMEOUT = 120 s, so the
socket timed out with EAGAIN while the refresh was running perfectly. In the GUI the timeout was
invisible: refresh_then ignores the reply and loads the page anyway, which re-read the old
timeline and made Refresh look like a dead button.
What changed. The daemon starts the refresh (spawn_timeline_refresh, which already dedupes)
and replies refreshing at once. Request::PhotosRefreshStatus reports whether one is running;
the gallery polls it every 2 s and reloads the page when it clears, and pdfs refresh photos
polls in the same way so the command still means "the timeline is up to date when I return".
refresh_timeline now logs photos, chunks and elapsed_ms.
What changed, second pass (2026-09-19). The refresh is incremental. photos.resolved_at
(schema v31) records that a photo's node was read, and refresh_timeline resolves only the photos
that have no such mark — the ones it has never read, and the ones a remote event marked stale.
A library that has already been read costs one enumerate_timeline and no enumerate_nodes call
at all.
The favourite question resolved by moving it off the refresh and onto the event feed:
classify_photo_events turns a NodeUpdated into photos_mark_unresolved(uid) instead of a
blanket "re-read everything", so a favourite toggled on a phone costs one node read rather than
63. A local toggle is claimed as our own echo (note_self_change), because the feed replays it and
a NodeUpdated carries nothing that says what changed. ContinuityLost / ScopeAccessLost and
pdfs refresh photos --full drop every mark, which is the one way back to a whole-library read.
TIMELINE_ENRICH_CHUNK is 150, matching the SDK's own MAX_BATCH_COUNT — a chunk of 200 was split
there into a full request and a 50-uid tail.
Still open. invalidate_albums() drives the same pattern per album
(crates/pdfs-fuse/src/albums.rs:96 and :191). It is bounded by album size rather than library
size, so it is nowhere near as expensive, but the technique that fixed this would apply there too.
B91 — The daemon hung overnight, and nothing could say what it was waiting on
Status: Open — the hang itself is undiagnosed. What is fixed (unverified) is the
undiagnosability: the daemon now reports what it is doing, warns when it stops making progress,
and is restarted by systemd when it stops answering.
Found: 2026-09-18, from a user report — the mount stopped responding and only came back after
systemctl --user restart.
Where: crates/pdfs-fuse/src/{workers.rs,diagnostics.rs,supervisor.rs,systemd.rs,mount.rs},
packaging/proton-drive.service.
Symptom. The mount stops answering. The process is alive and is not spinning; there is no panic, no error, and no log line after the moment it stopped. The accounting for the hung run:
Consumed 43min 39.045s CPU time over 20h 8min 15.733s wall clock time,
5.2G memory peak, 65.7M memory swap peak
It went silent at 17:10 and stayed that way until the manual stop at 13:18:54 the next day. A restart picked everything back up, which rules out corrupt local state.
What made it undiagnosable. Nothing in the daemon could be asked what it was doing:
- the worker pool published no state at all — not which threads were busy, not with what, not for how long, not how deep the queues were;
- control requests were not tracked, so a handler that never returned looked exactly like a handler that was never asked for;
- there was no watchdog, so a daemon that stopped answering stayed "active (running)" for twenty hours;
- resident size was never sampled, so the 5.2 G peak was only visible in systemd's post-mortem.
Related, and a real bug on its own. At 2026-09-18 09:58:35 a meta worker panicked during shutdown:
thread 'pdfs-fuse-meta-0' (3044) panicked at ...:
A Tokio 1.x context was found, but it is being shutdown.
The pool was flagged closed at teardown but never joined, so the process could drop the tokio
runtime while a worker was still inside a job. catch_unwind caught it, which is why it was only
ever a log line — but it is teardown ordering, not an error.
What changed.
Workerspublishes per-thread state (busy, job label, age) plus queue depths and per-lane completion counts, all outside the queue lock, and every FUSE handler labels the job it hands over.Request::Diagnostics/pdfs diagnosticsreports that, the in-flight control requests with their ages, the handler count against its limit, resident size and the pending-op count. It is built to answer while the daemon is wedged: atomics andtry_lockonly, with the one database read on a throwaway thread behind a 500 ms deadline.- A supervisor thread WARNs when a worker has held one job for 120 s, when a control request has run that long, or when a lane has a queue but completed nothing since the last tick; it samples RSS every 5 minutes.
- The supervisor answers the systemd watchdog, but only after a real round trip over the control
socket — so the ping means "a client can still talk to this daemon" rather than "a thread is
still scheduled".
WatchdogSec=120andRestart=alwaysturn the next hang into a restart instead of a morning of downtime, andMemoryHigh=2G/MemoryMax=6Gbound the growth. - Teardown joins the worker pool, with a 10 s deadline, before the caller drops the runtime.
Still open. None of this says why it hung. The candidates the evidence does not yet separate:
the 5.2 G working set (a leak, or a legitimately huge listing retained), a lock held across an
await somewhere in the control plane, and the ~150 rt.block_on sites on the control path that
have no op-level deadline of their own — each individual HTTP call is bounded by the SDK's 30 s
API / 300 s storage timeouts, but a sequence of them is not. The next occurrence should be a
one-line answer from pdfs diagnostics or from the stall WARN.
B90 — The local-file indexer walks into foreign FUSE mounts, and a dead one makes the daemon unkillable
Status: Fixed (unverified) — the walker no longer stats a nested mountpoint. The stuck-daemon
half cannot be fixed in userspace at all; see "Why there is no second line of defence".
Found: 2026-09-16, from a user report of systemctl --user restart failing in a loop.
Where: crates/pdfs-core/src/localindex.rs (default_excludes, scan),
crates/pdfs-fuse/src/background.rs (scan_local_once).
Symptom. Stopping the service times out, SIGKILL does not help, and every subsequent start
fails with another Proton Drive daemon is already using .../cache.db:
systemd[2196]: proton-drive.service: State 'final-sigterm' timed out. Killing.
systemd[2196]: proton-drive.service: Killing process 2221 (pdfs) with signal SIGKILL.
systemd[2196]: proton-drive.service: Processes still around after final SIGKILL. Entering failed mode.
systemd[2196]: proton-drive.service: Unit process 2221 (pdfs) remains running after unit stopped.
pdfs[73921]: ERROR mount failed; retrying in 5s error="open cache db: another Proton Drive daemon
is already using /home/narl/.local/state/proton-drive-linux/cache.db"
Diagnosis. The surviving process is a zombie that still has a live thread:
$ cat /proc/2221/status | grep -E 'State|Threads'
State: Z (zombie)
Threads: 2
$ cat /proc/2221/task/73231/comm; cat /proc/2221/task/73231/stack
pdfs-localindex
request_wait_answer
pdfs-localindex is parked in a FUSE request nobody will answer, in uninterruptible D state, so
it ignores SIGKILL. The thread group leader exits, the thread does not, the process cannot be
reaped — and it keeps the cache.db lock, which is what turns one stuck scan into a permanent
restart loop. The same dead connection also blocked an unrelated duf in fuse_statfs
(/sys/fs/fuse/connections/69/waiting was 2).
Cause. scan_local_once walks $HOME, and default_excludes only excluded the Drive
mountpoint plus our own state and cache dirs. The reporter's home also held
~/remote/narl.io (sshfs) and ~/remote/gdrive (rclone) — third-party FUSE mounts the walk
descends into.
scan does set same_file_system(true), which looks like it should already cover this. It does
not: in ignore 0.4.33 the check is is_same_file_system → device_num(path) →
path.metadata() (walk.rs:2164), i.e. a stat of the mountpoint, performed in run_one
after the directory has been queued. To discover that a mount is foreign, the walker first has
to touch it. Against a network FUSE mount whose server is gone, that stat never returns.
Fix. localindex::nested_mount_points reads /proc/self/mounts and returns every mountpoint
strictly below the scan root; scan_local_once appends those to the excludes. filter_entry runs
before the device stat (walk.rs:1926, before send), so a path-only rejection keeps the walker
from ever touching the mount. Mountpoint fields are octal-unescaped, because a \040 left in
place would silently fail to match and put the mount back in the walk. Re-read per scan, not
cached at startup: a mount that appeared since the daemon started is the likeliest to be a
half-alive network filesystem.
This changes nothing about what gets indexed. A nested mount is a different device, so
same_file_system(true) was always going to skip it — the exclude only removes the stat that had
to happen first.
Why there is no second line of defence. Once a thread is in uninterruptible FUSE wait, no
userspace change rescues it: not a join timeout, not std::process::exit, not SIGKILL. The kernel
will not reap the task until the syscall returns, which needs someone to answer the request or
abort the connection (echo 1 > /sys/fs/fuse/connections/<id>/abort, root-only). Avoiding entry is
the entire fix. That makes any new code path that stats a user-supplied or mount-crossing path
from a daemon thread a repeat of this bug — mirror-folder scanning is the obvious next candidate.
Recovery for an already-wedged daemon. As root, abort the connection with waiters
(grep . /sys/fs/fuse/connections/*/waiting finds it), which releases the thread, lets the zombie
be reaped and frees the cache.db lock. A reboot also clears it.
B89 — The recorded SDK dependency is three minor versions out of date in the working memory
Status: Closed (2026-10-06). proton-dev/CLAUDE.md no longer exists, so there is nothing left to correct. It was open as documentation only.
Found: 2026-08-17, while checking B85's evidence against the reader implementation.
Where: proton-dev/CLAUDE.md (the container-level file, not in this repo).
CLAUDE.md states that proton-drive-linux/Cargo.toml declares proton-sdk = "0.1" /
proton-drive-rs = "0.1" and that Cargo.lock pins 0.1.11 from the registry, against a local
proton-sdk-rs/ at 0.1.12. Neither half is true any more: the lock pins 0.6.0, bumped by
1.7.1 (see its changelog entry, "Workspace deps bumped to proton-sdk / proton-drive-rs 0.6.0")
and never reflected in the working memory.
Why it matters beyond tidiness. The whole point of that section is to tell you which source
you are reasoning about when you read SDK code. Verifying B85 meant establishing what read_at
does in the version the client actually builds against; following CLAUDE.md sends you to
proton-drive-rs-0.1.11 in the registry cache, and to a local checkout described as one patch
ahead of it when it is in fact five minor versions behind. The conclusion happened to be the same
in every version — but that was luck, and the next question of this shape may not be.
Fix: correct the versions, and prefer describing where to look (grep -A2 'name = "proton-drive-rs"' Cargo.lock) over restating a number that goes stale on every dependency bump.
B88 — pdfs pin, rename, move, delete and sharing still cannot name a path under a secondary mount
Status: Partly fixed. B86's fix covers ls and refresh. move takes paths in any location
through Core::move_between. rm, rename and mkdir were fixed and verified live 2026-10-02.
pin, unpin, opening a file, thumbnails, uploads, sharing and versions are still open.
Found: 2026-08-17, while fixing B86.
Where: crates/pdfs-fuse/src/control.rs, every handler still calling rel_to_mount.
B86's second half added Core::rooted_at and control.rs::route_to_mount, which resolve a path
against whichever mount the daemon owns it under and answer the request in that mount's inode
space. Only ListDir and RefreshScope::Dir were routed through it, because those two were what
the B86 repro needed: an escape hatch for a folder serving a stale listing.
Every other path-addressed control request — Pin, Unpin, Rename, Move, Delete,
CreateFolder, OpenFile, the sharing and public-link requests, ListRevisions — still uses the
primary-only rel_to_mount, so an absolute path inside a secondary on-demand mount is rejected
with "is not under the mountpoint". The user can now see such a folder's contents and cannot act
on them by path.
Why it was left. Routing a mutating request is not the same change as routing a read. Each of
those handlers resolves an inode, checks access, performs a remote operation and then invalidates
local state, and several of them invalidate self.state() — the primary — explicitly. Routing
them means auditing each for which mount's state it is really talking about, which is a larger and
riskier change than the one B86 needed. Doing it half-way would be worse than not doing it.
Required fix/test: move the remaining handlers onto route_to_mount, auditing each for
per-mount state assumptions as it goes; the for_each_state/uid-keyed helpers are the pattern for
the ones that touch more than their own inode space. Test: a pdfs pin and a pdfs rename by
absolute path inside a secondary on-demand mount, plus a path under no mount still erroring.
Fix so far (rm, rename, mkdir): the three handlers go through control.rs::route_change,
which is route_to_mount plus the mirror-folder hint and escape check that ResolveConflict
already had. The audit found nothing primary-only in Core::delete, Core::rename or
Core::create_folder. Core::rooted_at swaps in the fork's state, notifier, session flag and size
upgrades, the same per-mount fields Core::fork_state makes fresh. So self.state() in a routed
call already is the fork's inode space. My files and the on-demand folders do not overlap, so
delete forgetting the node in its own space alone is enough. rename already forgets it
everywhere (B74).
Routing raised a hazard the primary mount never had. An on-demand folder's root has a parent on
Drive, so pdfs rm ~/Downloads would have passed source_parent_uid and trashed the whole synced
folder from under its mount. rm and rename therefore go through route_node, which refuses a
path naming a mount's root. Test: only_a_path_below_a_mount_root_names_a_node. There is still no
handler-level test, because a Core can only be built against a live client.
Verified live (2026-10-02) in ~/Downloads, an on-demand folder. pdfs mkdir created a
folder. pdfs rename of a file whose upload was still queued reached Drive under the new name, at
6 bytes and with no conflict. pdfs rm trashed a file and then the folder. pdfs rm ~/Downloads,
pdfs rename ~/Downloads … and pdfs rm ~/ProtonDrive were refused as the root of a synced
location, and a path outside every mount as not under the mountpoint. mkdir and rm in
~/ProtonDrive still work. Not verified: the refusal in a mirrored folder, because the account has
none. The acceptance case "regression B88" repeats this on every location of a run, and an
--account run checks the refusal in the mirror folders it registers.
B87 — A revision's block geometry is learned a read too late
Status: Open — a bounded, self-correcting cost, accepted deliberately in 1.8.2.
Found: 2026-08-17, while fixing B85. This is the residue that fix leaves.
Where: crates/pdfs-fuse/src/reads.rs (Core::block_geometry, read_block).
B85's fix takes block boundaries from the revision instead of assuming 4 MiB. The geometry comes
from RevisionReader::block_sizes(), and a reader is expensive to open — link details, ancestor
keys, an S2K node-key unlock and the block table, which is exactly the per-read work the block
cache exists to avoid (B12). So the geometry is learned as a side effect of the first read,
which has to open a reader anyway, and written to a <key>.geom sidecar for later reads.
The first read of a file therefore still plans on BlockGeometry::uniform. On a file whose blocks
are not 4 MiB that plan is wrong, and it costs exactly one straddling fetch: the reader fetches
two of the revision's blocks to answer for one client block, and the block written to the cache is
at an offset the next read — now planning correctly — will not ask for, so it ages out unused.
Every read after the first is right.
Why it is not being fixed now. The alternatives are all worse than the cost. Resolving a reader before planning reintroduces B12's per-read key derivation on every file. Persisting the geometry at upload only covers files this client wrote, which are the ones the fallback is already correct for. Fetching the block table alone still costs the node-key unlock.
Worth doing if: non-4-MiB revisions turn out to be common on real accounts rather than a handful of files on one, or if the block table becomes reachable without unlocking the node key. Measure before deciding — one wasted fetch per file per revision is not obviously worth paying anything to remove.
B86 — An on-demand mount serves a listing that predates the mirror engine's own uploads
Status: Fixed in code (2026-08-17); live re-run of the repro pending.
Found: 2026-08-16, while trying to drive B42's remote-delete path from one machine.
Where: crates/pdfs-fuse/src/sync.rs upload path (no listing invalidation) meets the cached
children snapshot the on-demand mount enumerates from.
Repro. With folder 6 (~/pdfs-live-mirror) in mirror mode: create c1.txt and c2.txt,
pdfs sync now 6 — the activity feed reports c1.txt new file, c2.txt new file and
sync_entry gains rows carrying real remote_uids. Switch the folder to on-demand and list it:
$ ls ~/pdfs-live-mirror
b25.txt
b40.bin # c1.txt, c2.txt and the conflict copy are missing
ls ~/pdfs-live-mirror/c1.txt returns ENOENT. Restarting the daemon does not help — the stale
children snapshot is in the database, not in memory. Switching the folder back to mirror proves
the files were on the server the whole time: the pass downloads all five, and b40.bin comes back
md5-identical.
Effect. No data loss — the mirror side is correct and the files are intact remotely — but a
user who switches a synced folder to on-demand sees an incomplete folder, and anything that walks
it (a backup, an indexer, rm -r) sees the same incomplete folder. It also blocks B42's live
verification, because pdfs rm cannot name a file the mount will not admit exists.
Second, smaller gap in the same repro: pdfs refresh and pdfs ls reject any path that is
not under the primary mountpoint (Error: … is not under the mountpoint), so a secondary
on-demand mount has no manual escape hatch either.
Required fix/test: invalidate (or update) the parent's cached children when the sync engine
creates a remote node, the same way the FUSE write path does; and let refresh/ls resolve a
path under any mount the daemon owns, not just the primary. Test: the repro above, asserting the
new files appear immediately after the mode switch.
Fix (2026-08-17)
Two halves, matching the two halves of the report.
The stale listing. Core::invalidate_children_of (pdfs-fuse/src/lib.rs) drops a folder's
cached children by uid. The existing invalidate_parent_listing could not serve here for two
reasons: it names the folder by inode, and a mirror folder has no inode space at all at the moment
the engine writes to it; and the listing that outlived the pass was not a hot-cache map but the
listed flag on the folder's DB row — which is why restarting the daemon did not help. The new
helper clears that flag unconditionally and then invalidates whichever mounts happen to hold the
folder resident. The sync engine calls it after every remote mutation it makes: folder creation
and file upload (apply_one, upload_new), and the three remote-trash sites, whose parent uid
comes from the pass's folder-uid map or the baseline (parent_uid_of).
The missing escape hatch. StateRegistry now carries each mount's size_upgrades map as well
as its state, notifier and liveness flag — that set is exactly the per-mount half of a Core
(everything fork_state replaces) — and StateRegistry::covering_parts returns it for the mount
that most specifically covers a path. Core::rooted_at swaps those fields into a clone of the
Core, giving a view rooted at the owning mount plus the path relative to its root. pdfs ls
and pdfs refresh route through it (control.rs::route_to_mount), so a secondary on-demand mount
can be named at last; a path under nothing this daemon has mounted still gets the familiar "is not
under the mountpoint" error. The other path-addressed control requests deliberately still use the
primary-only rel_to_mount — routing the mutating ones is a larger change and was not what
blocked the repro.
Test: covering_parts_route_a_path_to_the_mount_that_owns_it covers the routing rule (nested
mount wins, suffix relative to it, its own inode space and upgrade batches, a path under no mount
routes nowhere). The listing-invalidation half is the live repro, which is what is still pending.
Test: the managed matrix of an --account run replays the repro. Before it switches a mirror
folder to on-demand, it writes pdfs-mode-switch.bin there and forces a pass. After the switch it
checks that the mount lists the file with its bytes, as "regression B86" under "mode switches".
Only the third pairing of the full matrix switches a folder that an earlier pairing had
on-demand, which is the stale case. --quick does not get there.
B85 — The read path assumes 4 MiB blocks; some revisions do not have them
Status: Fixed in code (2026-08-17) — the geometry now comes from the revision. See
"Correction to the diagnosis" below: the premise that this was losing data does not survive
reading the SDK, and the fix is a cost fix, not a correctness one.
Found: 2026-08-16, while live-verifying B84 (see its journal evidence)
Where: crates/pdfs-fuse/src/reads.rs, crates/pdfs-core/src/cache.rs (BLOCK_SIZE,
block_len)
Cause: The whole range path is built on one assumption — that a revision is a series of
BLOCK_SIZE (4 MiB) blocks with a remainder at the end. block_len(size, idx) encodes it, the
on-disk block cache is keyed by that index, and read_block asks the SDK's RevisionReader for
[idx * BLOCK_SIZE, block_len).
Some revisions on this account are not laid out that way, and B84's diagnostic caught it in the act. The warning it logs reads "shorter", but the numbers say the opposite:
bidx=0 blen=4194304 got=4939506 fsize=4939506 # one block holding the whole file
bidx=1 blen=745202 got=0 fsize=4939506 # …so there is no block 1
bidx=0 blen=4194304 got=5239804 fsize=11537412 # first block is ~5.0 MiB, not 4
bidx=1 blen=4194304 got=5670228 fsize=11106988 # …and they are not even uniform
bidx=2 blen=4194304 got=5233034 fsize=19932790
reader.size() agrees with fsize in every one of these, so B84's primary gate — the one that
catches a block table that does not add up — correctly does not fire. The file is intact and
the reader knows its real boundaries; it is this client that is asking for the wrong ones.
Why it is not currently losing data: B84's length check catches every one of these and routes
the file through the whole-file manifest-verified download, so reads are correct. The cost is
real, though: any file with non-4-MiB blocks is downloaded whole on first read, however few bytes
were asked for, and one above REPAIR_MAX (512 MiB) fails EIO and can only be read by pinning
it.
Fix direction: take the geometry from the revision instead of assuming it.
RevisionReader::block_sizes() already exposes it and B84 already logs it. That means block
offsets become a prefix sum over the real sizes rather than idx * BLOCK_SIZE, and — the part
that makes this its own change rather than a patch — the on-disk block cache is keyed and
length-validated on the 4 MiB assumption too (block_len), so its key scheme has to follow the
file's own geometry or the cache has to hold ranges rather than blocks.
Open question: what produced these revisions. Uniform-but-not-4-MiB would suggest another client's chunk size; these are not uniform (5239804, 5670228, 5233034), which looks more like a size recorded per block after some transformation. Worth answering before choosing the fix, since "blocks are whatever the revision says" and "blocks are a different constant" lead to different designs.
Correction to the diagnosis (2026-08-17)
The logged numbers cannot have come from read_at, and the "this client asks for the wrong
range" reading is wrong. RevisionReader::read_at(offset, length) does not assume any geometry:
it computes end = min(offset + length, file_size), plans over the real block sizes
(plan_blocks), and splices each block clamped to [offset, end) (splice_block). Both are
present and identical in the version the client actually builds against — Cargo.lock pins
proton-drive-rs 0.6.0, not the 0.1.11 the top-level CLAUDE.md still describes — and in
0.1.11 as well. So a read_at(bstart, blen) can return at most blen bytes, and every quoted
line has got > blen:
bidx=0 blen=4194304 got=4939506 # got exceeds the length requested
bidx=1 blen=4194304 got=5670228 # …as does this one
The only site in the client that logs this shape (reads.rs, the blen/got/fsize warning)
reaches it solely on the bytes.len() != blen branch of a read_at(bstart, blen). Those two
facts cannot both hold. Either the numbers were transcribed with blen and got swapped, or they
came from a build that is not in the history. This wants re-capturing from the journal before
anything is concluded from it — and in particular, the entry's claim that B84's repair path is
"currently masking data loss" here is not supported.
What is true, and is what got fixed: the client assumed the geometry. idx * BLOCK_SIZE and
block_len(size, idx) encoded "4 MiB blocks with a remainder" in the range planner, the in-memory
ring, the in-flight fetch key and the on-disk block cache. When a revision is not shaped that way,
nothing breaks — but every client block straddles two server blocks, so the reader fetches and
decrypts two to answer for one, and the block written to disk lines up with nothing the next read
asks for. That is a real and unbounded cost on exactly the files the entry found, and it defeats
the property the block cache was built around.
Fix (2026-08-17)
BlockGeometry (pdfs-core/src/cache.rs) is a prefix sum over a revision's block sizes, with
BlockGeometry::uniform(size) reproducing the old constant-block arithmetic exactly. Blocks are
now addressed by BlockSpan { idx, start, len } rather than by index, everywhere: the range
planner (spans(offset, end)), the in-memory ring, the in-flight-fetch key, and the on-disk block
cache. The span is the identity because the index alone stopped naming one byte range — block 1
under the fallback and block 1 under the real geometry are different bytes — so the block
sidecar records the block's plaintext start and a re-planned read misses rather than being
served the wrong offset. Sidecars written before that field existed read as "wherever the 4 MiB
assumption put it", which keeps the existing cache intact for every file this client uploaded.
The geometry itself is learned as a side effect of the first read (which has to open a
RevisionReader anyway) and written to a <key>.geom sidecar under the block cache;
Core::block_geometry reads it back and falls back to uniform when it is absent. Resolving a
reader just to plan would reintroduce the per-read key derivation the block cache exists to
avoid (B12), so the first read of a file still plans on the fallback and self-corrects. A recorded
table whose sizes do not sum to the node's size is refused outright, and eviction takes the
geometry with the blocks it describes.
The open question above is still open, and no longer blocks anything: "blocks are whatever the revision says" is now what the code does, so a different constant is just another table.
Tests: the_uniform_geometry_is_the_old_block_arithmetic,
a_non_uniform_geometry_places_blocks_by_prefix_sum (built on the exact sizes from the log
above), a_cached_block_is_not_served_to_a_different_geometry,
a_recorded_geometry_is_returned_only_for_the_revision_it_describes.
B84 — A block read that comes back short truncates the file, silently
Status: Fixed and live-verified 2026-08-16. Data correctness — an application reading a large
file through the mount could get a truncated copy that reported success. The truncation is now
impossible: a disagreement is repaired from the whole-file download, or the read fails EIO.
Re-verified on 1.8.1 after the release build: 118454731_p0.png, 121677947_p0.png,
123349685_p0.png and 128451683_p0.png each delivered exactly their stat size
(19471492, 11811519, 22055175, 11393178 bytes) with no Invalid argument reply in the journal.
Found: 2026-08-16, chasing what looked like log noise. GNOME's file indexer
(localsearch-3) was filling the journal with libpng error: Read Error,
libpng error: IDAT: CRC error and File too small to be a PNG on files under
~/ProtonDrive/Pictures, interleaved with 49 ERROR fuser::reply: Failed to send FUSE reply: Invalid argument (os error 22) from the daemon in one afternoon. The indexer was right and we
were wrong.
Where: Core::read_block (crates/pdfs-fuse/src/reads.rs:566), the assembly loop in
Core::read_range_remote (reads.rs:759-769); upstream, RevisionReader::read_at /
splice_block and rank_block_sizes (proton-drive-rs 0.6.0, revision.rs,
transport.rs:256).
Repro. Deterministic per file, no account state needed beyond the files themselves:
stat -c %s FILE # what the mount reports
cat FILE | wc -c # what the mount can actually deliver
| file | nodes.size / stat |
bytes readable |
|---|---|---|
Pictures/old/92433593_p0.png |
4317497 | 4194304 (tail block yields 0) |
Pictures/old/79779657_p0.png |
11574982 | 10567052 |
Pictures/old/2022-07-09_04-53.png |
6388959 | 4389310 |
pdfs pin on the third file, then re-read: 6388959 bytes, a fully valid PNG — every chunk
CRC passes, IEND is reached, the IDAT stream inflates to 44.8 MB. So the remote data is intact,
nodes.size is right, and the whole-file (manifest-verified) download path is right. Only the
block/range path truncates. Pinning is the workaround.
Root cause. read_block asks the SDK for blen = BLOCK_SIZE.min(fsize - bstart) bytes and
never checks how many it got:
let bytes = reader.read_at(bstart, blen).await.map_err(...)?; // reads.rs:577
RevisionReader::read_at clamps the range to its own file_size — summed from the believed
per-block sizes rank_block_sizes resolved — and splice_block clamps again to
plaintext.len(). Both are silent. A short block then vanishes in the assembly loop: with
block.len() short, e = (end.min(bstart + block.len()) - bstart) collapses, and for an empty
block s < e is false, so nothing is appended at all. The FUSE read returns fewer bytes than
asked for, the kernel hands userspace a short read, and every tool reads a short read as EOF.
The EINVAL reply errors are the same mismatch seen from the other side.
Nothing on this path validates a block against the length the file's own size implies
(block_len(fsize, bidx) — which the cache does check, in ContentCache::cached_block, and
which is why a short blob on disk is rejected and refetched while a short blob from the network
is served).
Not yet known: which rung of rank_block_sizes these revisions land on. The three
reader-vs-node deltas fit no constant ratio, so this is per-revision metadata rather than block
padding — the terminal block_count <= 1 fallback (vec![4 MiB; block_count]) would explain
the file that stops at exactly 4 MiB but not the other two. Needs a live probe printing
RevisionReader::block_sizes() against nodes.size for these three uids. That answer decides
whether the real fix belongs in this repo or in proton-sdk-rs.
Fix (2026-08-16). Core::read_block no longer trusts the range path blindly.
- The disagreement is caught before any bytes are read. If
reader.size() != fsize, the revision's block table does not describe this file and no block of it can be trusted — not just the short ones. Sizes that overstate a block shift every block after it and return full-length reads of the wrong bytes, which no length check would ever catch (the SDK'srank_block_sizesdocuments exactly this failure). So the whole file leaves the range path. - A short block is caught as a second net, compared against
block_len(fsize, bidx)— nowpubinpdfs-core::cache, because it is the contract the read path shares with the cache rather than a detail of the cache. The assembly loop inread_range_remotere-checks the same thing and failsEIOrather than contributing fewer bytes than the range it covers, which is the line that actually turned a bad block into a truncated file. - A file that fails either check is repaired, not just refused.
Core::repair_blockpulls the whole revision throughdownload_file_to— the manifest-verified path, which does no boundary arithmetic and is why pinning worked — into a scratch file, checks it againstnodes.size, and adopts it as the cached blob viastore_file. Every later read of the file is then answered byContentCache::read_rangebefore the block path is consulted at all. One download per file, not per block:content_repairsholds a lock per file so the concurrent block reads that all discover the same bad table share it. Bounded byREPAIR_MAX(512 MiB) — past that one unlucky read would pull down gigabytes, so it failsEIOand says to pin.
Nothing short is cached or promoted: read_block only reaches store_block and the block_ring
insert with bytes that passed. store_block itself deliberately does not check — its fixtures
store 4 KiB "blocks" of notional 1 GiB files to keep the budget tests cheap, and the read side
(cached_block, has_block) already rejects a wrong-length block as a miss. Validation is the
reader's job, at the point of fetch.
getattr needs no change after all: pdfs pin proved nodes.size is the correct one and the
reader's belief is the wrong one, so the fix is to stop serving the reader's belief — not to
publish it through stat.
Verified live (2026-08-16). Four multi-block PNGs under Pictures/wallpaper/ were read whole
through the mount on the running 1.8.1 daemon; every one delivered exactly its stat size
(5558846, 1098729, 2015463, 2291122 bytes), and no fuser::reply … Invalid argument appeared. The
journal shows the second net firing and the repair path serving each of them:
WARN block came back shorter than the file's size implies bidx=0 blen=4194304 got=4939506 …
INFO served a short block from a repaired whole-file download bidx=0 …
Answered, and not as expected. The open question was which rung of rank_block_sizes these
revisions take. The evidence says the reader is not the confused party: reader.size() equals
fsize in every case — the gate in step 1 never fires — and the blocks it returns are larger
than 4 MiB, not smaller (4939506 for a whole 4.9 MB file; 5239804 and 5670228 as first blocks).
This client's fixed 4 MiB block geometry is the thing that is wrong. Filed as B85; the repair
path above is what keeps those files readable in the meantime, at the cost of a whole-file
download on first read.
B83 — An access-deferred op retries forever and is reported as an ordinary queued upload
Status: Fixed (partly verified) 2026-08-16 — both halves are in and
unit-tested. The reporting half was confirmed against the live 28-op backlog on
1.8.0 (schema migrated to v27, all 28 rows window-stamped, WARN lines per op);
the resolution half has not yet been observed draining that backlog.
Found: 2026-08-16, running scripts/fuse-acceptance.sh --live on a
throwaway CLI-made mount. Four regression cases (B69, B70, B74, B79) failed with
TimeoutError: daemon mutation queue did not drain — the suite's
wait_for_queue() polls the global status.mount.pending_uploads, which had
been sitting at 28 for 30 days on the development machine. pdfs status called
them 28 queued uploads holding 18.4 GiB; pdfs transfers showed nothing;
failing_ops and parked_uploads were both 0.
Where: run_pending_drain / run_authorized_drain
(crates/pdfs-fuse/src/drain.rs), require_uid_writable
(crates/pdfs-fuse/src/lib.rs), Db::effective_node_access
(crates/pdfs-core/src/db/share_access.rs),
Db::defer_op_without_attempt (crates/pdfs-core/src/db/ops.rs).
Root cause. effective_node_access returns Ok(None) when the uid has no
row in nodes, and require_uid_access maps that to EACCES — "I have never
heard of this node" and "you may not write this node" are the same answer. The
drain reads EACCES as DrainDisposition::AccessDeferred and calls
defer_op_without_attempt, which by design touches neither attempts nor
last_error. So the op was re-deferred every DRAIN_ACCESS_RECHECK (5s)
indefinitely, and every channel that could have surfaced it stayed empty:
attempts 0, last_error NULL, parked counts only PARK_UNTIL rows,
failing counts only attempts >= FAILING_ATTEMPTS, and the one log line was
debug!. On the machine that found it, all 28 uids were absent from nodes:
select (select count(*) from pending_op p
where exists(select 1 from nodes n where n.uid = p.uid)) as in_nodes,
(select count(*) from pending_op) as total;
-- 0 | 28
Fix, part 1 — stop guessing. WriteAuthority { Writable, Denied, Unknown }
replaces the Result<(), Errno> the drain's access check used to return
(Core::uid_write_authority). require_uid_writable still collapses both
refusals to EACCES, which is right for a syscall: a stale handle must not be
admitted because the tree forgot the node. The drain does not collapse them —
Unknown becomes DrainDisposition::AuthorityUnknown(uid), and
Core::resolve_unknown_authority asks the remote:
- node still there →
upsert_nodere-interns it and the op is deferred for one recheck, so the next pass gets a real access answer; - node gone (
Ok(None)/is_gone) →record_op_failureimmediately, with a message naming the vanished uid, rather than waiting out the 5-minute window to say something vaguer; - could not ask → ordinary deferral.
The authority asked about is the op's own (pending_op_authorities), so a
create/mkdir refetches its parent, not the node it has not made yet. The
row and its staged blob survive every branch.
Fix, part 2 — say something. Schema v27 adds pending_op.access_deferred_since.
Db::defer_op_for_access stamps it on the first deferral of a run and returns
the stamp on every one; Core::defer_for_access warns on the first deferral and,
once the run passes DRAIN_ACCESS_DEFER_LIMIT (5 min), reports it through
record_op_failure so it enters the ordinary backoff and shows up in
failing_ops and the status error text. record_op_failure clears the window so
each escalation re-arms it rather than firing every recheck, and an attempt the
access check admits clears it too (clear_op_access_deferral). The row and its
staged blob are untouched on every path — this changes only what the user can
see.
The deferral window stays as the backstop for everything that does not resolve,
including a refetch that keeps not helping, and AccessDeferral now carries the
reason into both the log line and the recorded last_error.
Still open: the acceptance runner waits on the global queue
(ManagedSyncPair.wait_for_queue), so any unrelated stuck op fails those four
cases on a machine that has one — that is what turned this backlog into four
apparent FUSE regressions.
B82 — A parent_uid cycle hangs the daemon on any short search
Status: Fixed (unverified) 2026-08-12 — unit-tested, not driven against a
real corrupted database (nobody has one).
Found: 2026-08-12, writing a test for B80. The test built a two-node cycle
and called search; cargo test -p pdfs-core then sat at
search_excludes_a_parent_cycle for 29 minutes with zero CPU before it was
killed. Two earlier runs had already been abandoned to the same stall without
the cause being understood — it looks exactly like a cargo lock contention
problem, which is what it was first blamed on.
Where: path_of (crates/pdfs-core/src/db/utils.rs), upsert_node_tx's
descendants CTE (db/nodes.rs).
Root cause. Both walks are recursive CTEs using UNION ALL, which does not
deduplicate, so a.parent = b, b.parent = a never terminates. The index side
was never exposed — node_is_indexable_tx rejects a cycle — but a query shorter
than TRIGRAM_MIN (3 chars) takes the LIKE lane, and that lane reads
nodes directly and never consults the index. Every row it returns is then
handed to path_of. So typing two characters into the prompt was enough, and
the spin happens while holding the daemon's only SQLite connection: the mount,
the drain, and every control request stop with it.
The API is what would produce such a cycle (a rename/move race, a malformed event), so this is not a "corrupt your own DB" scenario only.
Fix. path_of is depth-capped at 256 and returns the truncated path rather
than hanging; path_relative_to already carried a 1024 cap, which is the
precedent. The descendants CTE in upsert_node_tx got the same guard for the
write side. Cycles also stay out of the index (see B80).
Still open: pin_is_pinned's ancestor walk (db/pins.rs) has the same
shape and no guard. Not reached from search, so it was left alone.
B81 — Prompt labels every Drive hit "My files", including device folders
Status: Fixed (unverified) 2026-08-12 — needs a GUI/dmenu run.
Found: 2026-08-12, reading the prompt while fixing B80: with B80 fixed, a
device-folder hit would render as My files / Downloads, a folder that does not
exist under the primary mount, while activating the same row opens
~/Downloads/…. Label and action disagreed.
Where: Hit::location() (crates/pdfs-gui/src/prompt.rs), shared by the
GTK prompt and the dmenu front-end.
The label was built from SearchHit::path with a hardcoded My files / prefix.
The daemon already resolves a device/sync-folder hit to a real local path in
mounted_path (Core::mounted_search_path), which is also what activation
uses; the label now derives from that when it is set, and falls back to
My files only for the primary mount.
Second defect, same area. Pin rows set is_dir from Pin::recursive
(prompt.rs, dmenu.rs). That flag is pin policy — whether the subtree is
kept on disk — not node kind, so a folder pinned non-recursively drew a file
icon and sent an invalid OpenFile. Pin now carries an is_dir: Option<bool>
resolved from the nodes row by pin_list, serde-defaulted so an older daemon
(which omits it) still parses; front-ends fall back to recursive only when the
node is not cached.
B80 — 29% of the account is absent from the search index
Status: Fixed (unverified) 2026-08-12 — migration replayed against a
read-only copy of the live DB; not yet run by the daemon itself.
Found: 2026-08-12, auditing cache.db after a user report that search felt
unreliable. pdfs search tickets returned only the My Files folder
Documents/Tickets, never the tickets.pdf sitting in the device folder
Downloads. Counting the tables explained it:
nodes 9252
nodes_fts 6547 ← 2705 rows never indexed
nodes in orphan subtrees 2705 ← exactly the missing set
Where: node_is_indexable_tx (crates/pdfs-core/src/db/nodes.rs).
Root cause — the same one as B79. A node was indexed only if walking
parent_uid reached a row with parent_uid IS NULL. Exactly one row is stored
that way (the My Files root). A device folder's root lives in device /
sync_folder and is never persisted as a nodes row, so Downloads,
Pictures, Documents, Music and Videos walk up to a parent that does not
exist and were judged unindexable, along with everything beneath them. Stale
pdfs-acceptance-* roots hit the same path.
The v16 backfill did not share the bug — it rooted its walk at any node whose parent is null or absent. So the index was correct immediately after migrating and then lost each subtree the first time one of its nodes was rewritten. That divergence between backfill and write path is why this survived: a fresh install looks fine.
Fix. Indexability now requires only that the walk terminates with nothing
trashed on the way; a cycle is still rejected explicitly (B82). MIGRATION_V21
(SCHEMA_VERSION → 21) rebuilds nodes_fts so existing installs recover
without waiting for each node to be rewritten, and fixes two gaps in the v16
walk it replays: descendants of a trashed folder stay out, and the walk is
depth-capped.
Replayed against a .backup of the live DB: 9276 non-trashed nodes → 9276
indexed (from 6547), 300 rows under Downloads, tickets.pdf at
Downloads/tickets.pdf.
Watch for. That path is the second thing this uncovered: the backfill first
seeded an uncached-parent root with an empty path, so the migration indexed
tickets.pdf while a later upsert indexed Downloads/tickets.pdf — the same
node with two paths depending on which code wrote it. Any future rebuild of this
index has to match path_of, not invent its own root rule.
Related, not a defect. The local half of the prompt indexes 110,187 files,
63,399 of them (58%) the Go module cache. Candidates are capped at 1000 and
scored after SQLite, so for a common term the pool filled with vendored files
before a real one was considered — for "test", 367 of the 500 best-ranked
candidates came from go/pkg/mod. SKIP_DIRS gained dist/build, a new
SKIP_PATH_SUFFIXES covers go/pkg/mod, and both candidate functions gained a
whole-query prefix lane ahead of the existing single-character one.
Test: the acceptance case "regression B80" writes a file with a unique token in its name,
rewrites it, and waits for pdfs search to find it. In an on-demand sync folder, that file sits in
the device-folder subtree the index used to lose.
B79 — Every write in an on-demand device-folder mount fails EACCES
Status: Fixed (verified live 2026-07-31, running in a locally rebuilt 1.2.1
package since)
Found: 2026-07-31, user on packaged 1.2.1: touch ~/Documents/test →
Permission denied, while the same operation under ~/ProtonDrive succeeds.
Where: crates/pdfs-fuse/src/state.rs, State::hydrate_access
Not a mode-bit problem — the mount root reads drwxr-xr-x and the kernel's
default_permissions check passes. The denial comes from the queue guard:
serve_create → Core::require_uid_writable, which intersects the access
every live inode space reports for that uid (lib.rs:721) before anything
reaches pending_op.
The on-demand fork itself says Owner. The primary state says Unknown:
Core::hydrate(lib.rs:1012) loads everynodesrow into the primary state, including the subtrees that belong to on-demand device folders — they are on the same volume, in the same table.- A device folder's
parent_uidis the device root, which is never persisted as a node. In this account's DB all six device-folder roots (Downloads,Pictures,Music,Documents,Videos,narl) haveparent_uid = <device root>with no matching row, so pass 2 parks them atORPHAN_INO. hydrate_accessthen hit its "parent is not resident" branch, which had no answer and fell back toAccess::Unknown— fail closed.Unknownis not writable, so the intersection denies every create, mkdir, write, rename, setattr and trash in every on-demand mount, permanently: those entries are never re-interned in the primary state, so nothing recomputes them.
~/ProtonDrive was unaffected because its nodes descend from ROOT_INO, which
ProtonFs::new hardcodes to Owner.
Two facts make this a classification bug rather than a policy one: the
persisted authority Db::effective_node_access already answers Owner for
exactly these uids (no share_access row covers them — the table holds only
virtual~sharedwithme and one accepted share), and the design table in
mount-architecture.md §2.2 says not under a share, no role → Owner, fail
open. The in-memory path disagreed with both.
Fix: the non-resident-parent branch now mirrors the persisted authority — a
recorded share_access row above the node still wins, otherwise fail open
to Owner on the mount's own volume and fail closed to Unknown on a
foreign one (State::is_own_volume). Foreign-volume shared content keeps its
fail-closed behavior, which is what B34 needs.
Second-order effect, also fixed by the same change: downgrade_known_shared_access
selects roots with entry.access != Access::Owner, so a ScopeAccessLost /
SharedWithMeUpdated event would additionally have force-downgraded every
device-folder tree to Viewer in memory.
Tests: state::tests::own_volume_nodes_without_a_resident_parent_stay_owned,
foreign_volume_nodes_without_a_resident_parent_fail_closed,
persisted_authority_outranks_the_own_volume_fail_open. The first fails on the
pre-fix branch with left: Unknown, right: Owner. The standing live guard is the
regression B79 acceptance case, which finds a mounted on-demand device folder
through pdfs locations --json and writes into it (skipping when this machine
has none).
Verified live (2026-07-31): the packaged 1.2.1 daemon denies
touch ~/Documents/pdfs-b79-test, mkdir, and > redirect while
~/ProtonDrive accepts them; a local --release build of this fix accepts all
three on the same account and DB, the write drains (pending_op back to 0, the
new nodes carry real G88km… remote uids rather than local ones), and the
daemon log has no 403/EACCES. ~/ProtonDrive/Shared with me/ stayed
dr-xr-xr-x, and the accepted editor-role share under it stayed writable — the
foreign-volume fail-closed path is unchanged.
Side note for whoever runs this next: stopping the daemon and touching a
file in an on-demand mountpoint while it is unmounted leaves that file in the
local directory, and restore_ondemand_mounts then refuses to mount over the
non-empty dir (WARN … local dir is not empty; refusing to mount over it).
Empty the directory and restart the unit.
B78 — Profile backup fails: Cannot create file at the root of a device
Status: Fixed (verified live 2026-07-31). Same defect as B68, which was
found first on the managed-live matrix; B68's second half — making backup health
visible — is still open.
Found: 2026-07-29, user reported upload profile: proton api error NotEnoughPermissions (http 422): Cannot create file at the root of a device
The background task that backs up the machine's profile (sync folder mappings, pins, cache budget) currently attempts to write profile.json directly to the device root node (device.root_uid) on the Drive backend. The Proton Drive API has started enforcing a restriction that files cannot be created directly at the root of a device; they must be placed inside a folder.
Consequence: The upload is rejected with HTTP 422. The daemon retries when changes happen but the profile is never successfully backed up. Local changes (pinning, adding a sync folder) still take effect locally, but they will not be restorable on a new machine.
Fix applied (matches the original RECOVERY.md spec): the profile now lives
in a .proton-drive-linux folder inside the device root.
pdfs_core::profile::PROFILE_DIR_NAME— the new constant, next toPROFILE_FILE_NAME.profile.rs:ensure_profile_dir— reuses the folder by name via the existingfind_device_child_folder, creating it on the first save. Reuse-by-identity matters: a second folder of the same name would split the record in two with nothing to say which is current.profile.rs:save_profile— both upload branches (new file, new revision) now address the folder, never the device root.profile.rs:load_profile— reads the folder first, then falls back to a device-rootprofile.jsonwritten by an older client. The fallback is read-only and the legacy document is not trashed: another machine may still be running a pre-fix client that reads it.profile.rs:list_restorable_folders— filters.proton-drive-linuxout, or the restore picker would offer to sync the bookkeeping back onto the machine it describes.devices.rs:add_sync_folder— refuses a local folder named.proton-drive-linux, which would otherwise be reused as the remote profile folder by name and then reconciled (i.e. overwritten) by the sync engine.
Verified live (2026-07-31) against the real account: first save logged
created device profile folder name=".proton-drive-linux" then
profile backed up folders=6 with no 422; a second trigger reused the folder
(no second create) and uploaded a revision; pdfs sync restore --json lists
exactly the six device folders and not the profile folder.
Tests: profile::tests::the_profile_is_never_uploaded_to_the_device_root,
loading_falls_back_to_a_legacy_device_root_profile,
the_profile_folder_is_not_offered_as_user_data — source-level guards in the
style of lib.rs's queue-guard ordering tests, because both upload branches are
pure network calls with nothing to assert offline.
Not covered by this fix: B68 also asks for backup health to be visible
(currently a WARN in the log and nothing in the UI), and for the
replacement-machine / fresh-state restore to be re-run end to end.
B75 — Network work on fuser's single dispatch thread freezes the whole mount
Status: Fixed in-tree 2026-07-26 — live-validated. Found 2026-07-26 while
chasing what was thought to be a B74 regression.
Impact: any burst of concurrent writes froze the entire mount — not the
files being written, the mount. ls on an unrelated directory of it timed out
for minutes. The daemon itself stayed healthy throughout: the primary mount and
the control socket answered normally while a sync-folder mount was unreachable.
Root cause. mount.rs and devices.rs both build their session from
Config::default(), which leaves n_threads unset, and fuser 0.17 defaults it
to 1 (session.rs:254). One dispatch thread per mount reads a request, runs the
handler to completion, and only then reads the next. Six handlers were handed to
the [Workers] pool — lookup, getattr, readdir, read, getxattr and the
slow serve_lookup — and everything else ran inline, including handlers that
block on the API:
| handler | blocking work |
|---|---|
create |
upload_file to mint the node, then fetch_node — two round trips |
mkdir |
create_folder, then fetch_node |
rename |
rename_node, then move_node, plus a trash on the replaced-destination path |
unlink / rmdir |
trash_nodes; rmdir enumerates the folder first |
setattr (path truncate) |
queue_truncate → fill_gaps, which reads the kept prefix from the remote |
So eight applications each creating a file served one create at a time, and
every read, open and lookup on that mount queued behind them.
Measured, acceptance suite's independent and shared-file concurrency (8
threads × 16 files), and a 3-second ls of the mount root sampled every 4 s
throughout:
| before | after | |
|---|---|---|
| concurrency case | 200 s watchdog ×2, 37 s ×1 | 16.9 / 24.1 / 29.1 s, all pass |
ls of the mount during the run |
33 of 55 probes timed out | 0 of 85 |
readdir stability |
148.8 s | 62.9 s |
namespace operations |
31.3 s | 15.7 s |
Fix: each of those handlers now parses its arguments inline — cheap, and it
keeps the error paths prompt — and hands the body to the worker pool as a
serve_* method, the pattern lookup/readdir already used. Metadata work
takes Lane::Meta; the truncate path takes Lane::Transfer because it can pull
a whole file.
release was deliberately left inline, and this is load-bearing. Moving it
too made all three concurrency runs fail with concurrent file 1 mismatch. The
kernel does not wait for a release reply before letting close(2) return, so
the only thing sequencing the staging of the written bytes against the open +
read an application issues immediately afterwards is that both are served by
the same dispatch loop. From a worker, the read overtakes the staging and is
answered from the remote's older revision. Taking release off the loop — worth
doing, since queue_revision gap-fills a partial write from the network — needs
a per-node "staging in flight" barrier that reads wait on, not the dispatch
loop's accidental ordering.
Not fixed here. n_threads is still 1; with the blocking handlers moved off,
the loop is short enough that raising it is a separate question. Related findings
from the same audit, still open: cache/reader validity is keyed on
(mtime, size) while the revision identity is active_revision_id
(cache.rs:84 — two same-size revisions within one mtime second collide, which
is what a SQLite page rewrite looks like); local_finish_scan holds the daemon's
one SQLite connection across a full FTS5 trigram rebuild (db/local.rs:79);
earliest_due_at has no <= now filter where next_due_op does, so the drain
busy-spins when offline with a due op (db/ops.rs:385 vs :421); and
drain_local_node deletes its op before the fallible adopt_real_uid
(drain.rs:346).
B74 — A drained create never reaches a sync-folder mount's inode space, so the file reads as empty
Status: Fixed in-tree 2026-07-26 — live-validated. Found 2026-07-26 by the
expanded acceptance suite (regression B70 case), then isolated to a broader
trigger.
Impact: on a sync folder mount (not ~/ProtonDrive), a file whose create
was still queued reads back as 0 bytes for as long as the daemon runs, even
though the bytes are on Drive intact. This is the write-to-temp-then-rename
pattern used by every browser download, most editor saves, rsync, and git.
Not data loss. An earlier revision of this entry called it permanent data
loss; that was wrong and is corrected here. The uploaded revision is complete and
correct: for every "lost" file the node row carried the right size and
active_revision_state: "Active", and a daemon restart made all of them read
back with the expected sha256. Nothing had to be recovered. The defect was
entirely local and entirely in the read path.
Reproduction (on an on-demand mount, per file: write, close, rename, wait for
pdfs status to report the queue drained, then read back):
| scenario | result |
|---|---|
| plain name, never renamed | content intact |
| plain name, renamed after close | 0 bytes until restart |
.crdownload, renamed after close |
0 bytes until restart |
.crdownload, never renamed |
content intact |
| already-settled file, renamed | content intact |
| any of the above, after a daemon restart | content intact |
The gap between close() and rename() does not matter — 0 s, 2 s and 20 s all
behave the same. The boundary is not timing but state: the affected node is one
whose create op was still queued, on a mount that is not the primary one.
Root cause — one drain thread, many inode spaces. The daemon mounts
~/ProtonDrive plus one FUSE session per ondemand sync folder. Each of those
is a Core::fork_state clone with its own State — its own entries,
by_uid, children, next_ino — while sharing db, cache, pending and
client. There is exactly one pdfs-drain thread in the process, owned by
the primary mount, and it serves the whole shared pending_op table.
So when a create queued by a sync-folder mount lands, the drain calls
adopt_real_uid — which looked only at self.state, the primary mount's inode
space. The fork's entry keeps its local~… placeholder uid forever. And
Core::read_range (crates/pdfs-fuse/src/reads.rs:314) short-circuits a local
uid to an empty vec:
if is_local_uid(uid) { return Ok(Vec::new()); }
which is why the read is silent, instant, and zero-length. A restart rebuilds
every state from the DB, where remap_local_uid had already written the real
uid, so the file reads correctly from then on.
Confirmed live: read handler uid=local~1785067300752-0 fsize=140000 three
seconds after pending create landed, and at the same moment
adopt_real_uid … hit=None strays=[] — the primary state had no entry for the
uid at all. /proc/<pid>/task/*/comm showed 1 pdfs-drain against 8 fuser
sessions.
Fix: Core gained a states: Arc<StateRegistry> — a Vec of weakly-held
mounts that every mount publishes itself into (Core::register_state, called
from mount.rs for the primary and from fork_state for each fork), plus
Core::for_each_state, which walks every live inode space one lock at a time.
A uid is unique across mounts, so at most one state matches and the rest are
no-ops. Five drain sites were converted: adopt_real_uid, adopt_drained_name,
the conflict-copy invalidate_listing, the active_writes baseline rebase in
refresh_after_upload, and that function's closing intern.
It was completely silent. No ERROR, no WARN, no failed op, no conflict
copy, no retry — pdfs status reported the queue drained, because it had. Any
diagnosis that relies on the journal or on queue state will not see this.
Regression cover: scripts/fuse-acceptance.py, case
regression B74: rename after close keeps its content — the narrowest form
(plain names, no overwrite, no transient suffix). It failed in two consecutive
live runs before the fix and passes after it. The pre-existing B70 case failed on
the same cause and now passes too.
Audit of the rest of the daemon. Every background-thread state.lock() was
reviewed for the same shape. The rule that came out of it: a uid is unique
across mounts, so uid-keyed work must be broadcast; an inode number is only
meaningful to the mount that minted it, so inode-keyed work must not be. What
changed:
| file | site | why |
|---|---|---|
background.rs |
apply_event |
the read-side twin of this bug — remote deletes, trashes and renames only ever reached the primary mount's tree |
sweep.rs |
conflict-copy forget_or_unlink |
a conflict copy inside a sync folder lives in that fork's inode space |
sweep.rs |
is_busy |
answered "idle" for every file open on a fork, letting the sweep delete one with a writer attached |
sharing.rs |
rel_path_for_uid |
a share of a node in a sync folder could not be named |
drain.rs |
node_place |
fell through to a DB lookup for fork-resident nodes |
lib.rs |
queue_trash, rename, remove_replaced (×2), drop_local, restore |
uid-keyed forget/invalidate |
Deliberately left on self.state: upload.rs (parent_ino/pino are inode
numbers), the size-upgrade path in lib.rs, and drain.rs's root_uid, which
is correctly primary-only. The 38 filesystem.rs sites are FUSE callbacks —
they already run on the right mount's session thread. parent_is_gone never
touches state at all; an earlier note here saying it did was wrong.
The registry is now a StateRegistry type with a unit test
(state_registry_walks_every_live_mount_and_reaps_the_rest) covering the case
this bug was: a fork that fails to appear in the walk, or one that lingers after
unmount.
One regression came out of the audit, and is fixed. Broadcasting
apply_event's parent-listing invalidation meant forks started receiving
invalidations they had never received — and it invalidated a folder once per
event, while a burst of our own uploads produces one event per file all naming
the same folder. Every following lookup in that folder then re-enumerated it
from the API: quadratic in the folder's size.
Two mitigations, both in background.rs. A batch is collapsed to one
invalidation per folder — apply_event records the parent in DirtyParents and
flush_dirty_parents publishes it once after the batch loop (every break path
falls through to it). And a change the daemon made itself is recognised as its
own echo rather than treated as foreign: the drain records it
(Core::note_self_change, consumed once within SELF_CHANGE_TTL_MS), which
stops the feed's report of our own upload from evicting the content blob we just
sent. Neither weakens the result — the state after a batch is what it always was.
The hang that showed up alongside this was a different bug, and this entry
originally misattributed it. readdir stability at 133.94 s and a concurrency
case that blew past its timeout were blamed here on the invalidation storm. They
were not: the mount was freezing because network work runs on fuser's single
dispatch thread, which is B75. Fixing that took the concurrency case from two
timeouts in three runs to 16.9 / 24.1 / 29.1 s and readdir stability to 62.9 s.
The coalescing above is still worth having — it is a real amplification — but it
was never the stall.
Note on B70: B70's own fix was never implicated. The .crdownload park/un-park
behaves correctly in isolation; the empty read arrived with the rename, on the
same path a plain temp file takes. B70's live validation was blocked on this and
is now unblocked — its acceptance case passes.
B73 — Unimplemented FUSE handlers log a WARN per probe
Status: Fixed in-tree 2026-07-26 — live-validated. Found 2026-07-26 by the
expanded acceptance suite.
Found: scripts/fuse-acceptance.sh --live step "unsupported operations refuse
cleanly". Each probe produced a multi-line fuser warning in the daemon journal:
WARN fuser: [Not Implemented] symlink(parent: INodeNo(0x2d), link_name: "symlink", target: "link-target")
WARN fuser: [Not Implemented] link(ino: INodeNo(0x94), newparent: INodeNo(0x2d), newname: "hardlink")
WARN fuser: [Not Implemented] mknod(parent: INodeNo(0x2d), name: "fifo", mode: 4516, umask: 0x12, rdev: 0)
WARN fuser: [Not Implemented] setxattr(ino: INodeNo(0x91), name: "user.pdfs.probe", flags: 0x0, position: 0)
WARN fuser: [Not Implemented] lseek(ino: INodeNo(0x93), fh: 0, offset: 0, whence: 3)
Cause: ProtonFs does not implement these callbacks, so fuser's default
trait method answers ENOSYS and logs at WARN. The refusal is correct — the
acceptance suite accepts every one of them as a clean refusal — but the logging
is not conditional on anything the user did wrong.
Impact: cosmetic, but these are not exotic calls: git probes symlinks,
cp -a probes hard links, tar and cp --sparse probe SEEK_HOLE/SEEK_DATA
(whence: 3 above is SEEK_DATA), and any xattr-aware tool probes setxattr.
Warnings that look like faults bury real ones and make both --journal-check
and manual triage noisier.
Measured across two full acceptance runs against one daemon lifetime (PID
887441, mount never remounted):
| op | warnings | behaviour |
|---|---|---|
link |
11 | repeats on every call |
symlink |
4 | repeats on every call |
mknod |
2 | repeats on every call |
setxattr |
1 | kernel cached the ENOSYS; never asked again |
lseek |
1 | cached |
copy_file_range |
1 | cached |
fsyncdir |
1 | cached |
So the FUSE kernel module remembers ENOSYS for the file-level operations and
stops issuing them for the lifetime of the mount, but not for the namespace
operations, which are re-issued forever.
link leads for a sharper reason than "tools probe it": git uses link() to
place every loose object, falling back to rename() when it fails. In one
workloads run, 9 of 10 link warnings carried a 38-hex-character name — the
.git/objects/ab/<38 hex> layout — from just 3 commits over 3 files. The volume
is therefore proportional to work done, not a fixed startup cost: a git clone
of a real repository emits one multi-line warning per object. The lone symlink
that is not the acceptance probe is git's core.symlinks detection at
git init (a random name pointing at testing).
That moves this from cosmetic to a genuine operational problem: ordinary use of git inside a mount can bury every other log line.
Fix (applied): all seven callbacks are now implemented explicitly in
crates/pdfs-fuse/src/filesystem.rs, each answering with the same errno
fuser's default would and nothing else — no WARN. Following ioctl
(which already used this pattern and says why in its comment), the errno
contract is unchanged, so the acceptance suite's clean-refusal assertions hold
byte for byte:
| op | errno | why that one |
|---|---|---|
symlink, link |
EPERM |
meaningful operation, unrepresentable on Drive; git falls back to rename() |
mknod |
ENOSYS |
no devices/FIFOs/sockets; regular files arrive via create |
setxattr |
ENOSYS |
surfaces as EOPNOTSUPP; reads stay implemented for the thumbnail interface |
fsyncdir |
ENOSYS |
kernel absorbs it and reports success; dir changes are already durable in SQLite |
lseek, copy_file_range |
ENOSYS |
must stay ENOSYS — it is what makes the kernel emulate SEEK_HOLE/SEEK_DATA and coreutils fall back to read/write instead of failing |
Only the three namespace ops were strictly required (the other four silence
themselves once the kernel caches ENOSYS), but implementing all seven keeps the
refusal explicit and documented at the call site rather than inherited.
Verified: full live acceptance run after the fix — 19 pass / 1 skip, the
capability diff identical to the pre-fix run, and 0 Not Implemented lines
in the journal across the whole run (was 17, 10 of them link).
Regression cover: the acceptance suite's clean-refusal probes assert the
errno contract; they pass unchanged.
B72 — pdfs trash never returns
Status: Hardened in-tree 2026-07-26 — the hang is gone by construction; the
underlying slowness of the first refresh is still unmeasured (live-pending).
Found: capturing a trash baseline before the B69/B70 live validation. pdfs trash
produced no output and was still running when killed at 120 s. Daemon was healthy
and every other CLI call (pdfs sync list, pdfs --version) returned promptly.
Where: Core::list_trash (crates/pdfs-fuse/src/lib.rs). Not the CLI: the
client's 120 s READ_TIMEOUT (pdfs-core/src/control.rs) is what ends the call,
and the daemon never writes a response line.
Root cause. The trash of this account had never been fetched successfully —
trash table empty, no trash_synced_ms in sync_state. list_trash therefore
always took its "never fetched → rt.block_on(refresh_trash())" branch, which:
- materialized the entire trash in one
enumerate_nodescall (one S2K unlock per node) before persisting anything, so nothing was ever saved and no later request could be served from the DB; - had no single-flight guard on that branch (only
spawn_trash_refreshdid), so every attempt started another full refresh; - was never cancelled when the requester timed out, because the control handler
sits in
block_onon its own thread and nothing signals it.
So each look at the trash added a permanent, full-cost refresh to the daemon.
Measured on the live daemon after a handful of GUI/CLI attempts: 11 stale
pdfs-control threads, RSS 2.9 GB. Concurrent with a drain backlog that was
uploading conflict copies continuously (1991 Conflict activity rows, multi-GB
uploads), none of the refreshes ever finished.
Fix (in-tree). list_trash never blocks on the whole refresh: it kicks the
single-flighted background refresh, waits at most TRASH_FIRST_WAIT (20 s, well
under the front-end read timeout) on a Notify, and answers with whatever has
materialized. refresh_trash now materializes in TRASH_MATERIALIZE_CHUNK (150)
batches, persisting the cumulative listing after each one, and logs the uid count
plus per-batch and total elapsed at INFO/DEBUG — previously the whole path was
silent, which is why the stall could not be located from the journal.
Still open: why one refresh is slow enough to matter. The new INFO lines
(trash refresh: enumerated / batch / done) answer that on the next run:
whether enumerate_trash_node_uids returns at all, how many nodes the trash
holds, and the per-batch cost. If the count is large, materializing with
enumerate_nodes_light (skips the per-file node-key S2K; size would fall back
to on-storage size) is the next lever.
Impact: the trash listing is the user's recovery path after any destructive operation, including the B71 sweep. It needs to work before the sweep is allowed to trash anything.
B71 — Conflict sweep is not production-ready (auto-trash loop, CRIT-13)
Status: Blockers 1–4 fixed in-tree 2026-07-26; items 5–10 and the remaining
cleanup still open. The sweep is safe to leave running (it is report-only by
default and cannot trash anything until explicitly switched to enforce).
Where: crates/pdfs-fuse/src/sweep.rs, spawn site crates/pdfs-fuse/src/mount.rs,
config surface crates/pdfs-core/src/config.rs.
The B69 fix (see above) shipped a pdfs-conflict-sweep thread that trashes user
files on the real Drive from a background loop. The identity check itself is
conservative (equal size and equal content_sha1, missing digest treated as
divergent), but the machinery around it is not yet safe to run unattended. Review
found the following, ordered by severity. Each is a separate defect; they are
grouped because they gate the same feature.
Blockers (all fixed 2026-07-26 — the sweep may not trash anything without them):
- Data-loss race — trash-then-discard (CRIT).
remove_conflict_copy(sweep.rs:137-155) trashes the node remotely and then callsdiscard_queued_ops(uid), which deletes the node's queued ops and discards their staged blobs (lib.rs:1544-1553). The decision was made from theload_all()snapshot taken atsweep.rs:77, with network round trips in between. A user edit landing in that window is destroyed — the queued write is dropped and its staged bytes, which may be the only copy, are discarded. This violates the invariant thatstaging/is never purged. Fix: re-read the node from the DB and re-verifyactive_revision_idimmediately before trashing; skip the node outright ifCore.pendingor the op table has any entry for it; never let the sweep discard staged bytes. - No kill-switch (CRIT).
mount.rs:234-239spawns the loop unconditionally. There is noAppConfigfield and no environment override, so the only way to stop a destructive background loop is to uninstall. Fix: addAppConfig.conflict_sweep: Option<bool>(sameOptionidiom ascache_budget/ignore_patterns), aPDFS_CONFLICT_SWEEP=offoverride for support, and a Settings toggle. - Open handles ignored (HIGH). Nothing checks for a live file handle before
forget_or_unlink+evict_reader, so the sweep can trash a node a reader currently has open. Fix: skip nodes with open handles. - Enforcing on day one (HIGH, process). The sweep exists because B69 proved
the previous revision-identity check was wrong; betting deletion on the
replacement check in the same release, with no field evidence, is the wrong
order. Fix: ship report-only (log
would-trashplus the existingActivityKind::Conflict), collect a release cycle of real logs, then flip to enforcing.
Fix for 1–4 (landed 2026-07-26):
SweepMode(pdfs-core/src/config.rs) —Off/Report/Enforce, serialised intoAppConfig.conflict_sweep(#[serde(default)], so configs predating the field still load).DefaultisReport, which is blocker 4: an existing install that has never heard of the setting cannot get the enforcing behaviour on upgrade.AppConfig::resolved_conflict_sweeplayersPDFS_CONFLICT_SWEEPover the stored value; the precedence is a pureresolve_sweep_mode(env, stored)so it is testable without mutating process state. An unparseable value falls through to the config rather than guessing a mode in either direction.- Spawn gate (
mount.rs) —SweepMode::Offskips the thread entirely rather than starting an idle one, so the setting is verifiable from outside the process (ls /proc/<pid>/task/*/comm). Both branches log the resolved mode at startup.mountgrew past clippy's argument limit, sousername+sweep_modemoved into aMountOptionsstruct resolved by the caller —mountstill reads neither config nor environment. - Interlocks (
sweep.rs::duplicate_still_removable) — re-checks, immediately before the destructive call, everything the pass decided from its stale snapshot: no queued op (newDb::has_any_op, a cheapCOUNT— a queued op means bytes are owed an upload anddiscard_queued_opswould throw away the staged blob holding them), no open write handle or shared scratch file (Core::is_busy, checkingactive_writes+handlesunder oneStatelock), and the node still at the same revision id, size, digest, name, parent and untrashed state. Any doubt — including a failed DB read — means leave it alone: a copy that survives to the next pass costs nothing, a wrongly removed one costs data. This is blockers 1 and 3. - Testability — the decision moved into a pure
decide(node, sibling, base, mode) -> SweepAction, the same extraction used forrevision_changedandis_own_self_supersede; everything that can fail or race stays in the caller. Tests now cover removal-when-enforcing, report-mode-never-removes, size-differs, digest-differs, missing-digest in both directions, and orphaned copies. Gate:cargo fmt/clippy -D warningsclean,cargo test --workspace282 passed (was 270).
Should fix:
load_all()every pass (MED). Each 300 s pass pulls the entire node table into aVec<StoredNode>and builds a full(parent, name)HashMap. Fix: query only conflict-named rows (name LIKE '% (sync-conflict %') and look up siblings per parent — O(conflicts), not O(drive).- No batching or per-pass cap (MED). The B69 incident produced 151 copies;
that is 151 sequential
trash_nodescalls in one pass with no cap and no backoff.trash_nodestakes a slice — batch it and cap per pass. onlinesampled once per pass (MED). Read atsweep.rs:100, ahead of a loop that does network I/O; going offline mid-pass yields a long run of failing calls. Fix: bail on the first transport error.- Ambiguous sibling index (MED).
by_parent_name.insert(sweep.rs:97) is last-write-wins, so with duplicate(parent, name)rows the surviving entry depends onload_allrow order. Fix: make the choice deterministic or skip ambiguous parents. - Trashed ancestors (LOW). A copy under a trashed folder finds no live sibling and is flagged "orphaned", spamming the activity feed. Fix: skip nodes with a trashed ancestor.
- SHA-1 provenance unverified (MED).
is_identical(sweep.rs:181) compares digests without asserting they belong to the currentactive_revision_id. A staleXAttrdigest from an earlier revision plus a size collision would delete real data. (The missing digest case is already handled correctly.)
Also required to close:
- Tests. The policy is now covered via the pure
decide(see above). Still untested are the interlocks, which need aCore: pending-op→skip, open-handle→skip, revision-moved→skip, offline→skip. Those want a test harness that can build aCoreagainst a temp DB and a stub client. conflict_notifiedis unbounded.HashSet<NodeUid>that only grows except on trash (sweep.rs:157,170).- Docs.
docs/ARCHITECTURE.mdthread map does not listpdfs-conflict-sweep;docs/RECOVERY.mdneeds a "the sweep trashed a file, recover it from Proton trash" path.
Mitigating context from the 2026-07-26 pre-run snapshot: the account had only
2 conflict copies left, both divergent in size and content_sha1, so the
correct behaviour for the first live pass is to trash nothing.
B70 — Browser in-flight temp files uploaded and conflict-forked on the on-demand mount
Status: Layers A + B fixed in-tree 2026-07-25 — live-validated 2026-07-26.
Validation was blocked until B74 was fixed (its scenario ends in a rename, which
B74 made read back as empty); with that out of the way the acceptance case
regression B70: transient download name is not sealed passes live.
Found: triaging the divergent conflict copies B69 left behind. In ~/Downloads
(an on-demand mount) the complete files had been forked into (sync-conflict)
copies while their truncated partials kept the canonical name — e.g.
teamspeak.tar.gz (30 MB partial live vs a 3.77 GB conflict copy) and nils.zip
(17 MB vs 3.18 GB). The conflict copies were byte-exact to the user's independent
gdrive backups, proving the conflict copy was the finished download and the
live file the stub. The smoking gun: four Unconfirmed NNNNN (sync-conflict …).crdownload
copies — Brave's in-flight temp files, forked mid-download. User confirmed the
trigger: downloading with Brave directly into the on-demand folder, where a stall
- "resume" is routine.
Where: on-demand write path —
crates/pdfs-fuse/src/filesystem.rsrelease(queue_revision, ~line 780) andrename(~line 933); drain seal incrates/pdfs-fuse/src/drain.rs. The existing ignore system (crates/pdfs-core/src/syncignore.rs,DEFAULT_IGNORE_PATTERNS) is wired only into mirror reconcile (sync.rs), so it never sees on-demand writes.
Cause: A browser downloading into the mount writes a growing temp file
(*.crdownload, or *.part for Firefox) and, on a stall, closes the fd. release
hands every closed write to queue_revision, so a partial revision gets sealed
as the canonical file. On resume the browser reopens and rewrites the full file;
its baseline revision has moved (the client sealed the partial itself), so the
drain diverts the finished bytes into a (sync-conflict) copy. Net: the stub wins
the name, the real file is exiled to a conflict copy, and every abandoned
.crdownload is uploaded as its own multi-hundred-MB node. Unlike B69 this is a
genuine two-revision divergence, so B69's revision-id detection correctly refuses
to auto-sweep it — the fix has to stop the fork from happening, not reconcile it
after.
Fix — two layers, both without an SDK release (A stops transient names ever sealing; B stops a single-writer resume forking against its own seal):
- A — transient names never seal a revision until finalized (done 2026-07-25).
A new
syncignore::is_transient_namerecognises browser download temps (*.crdownload/*.part/*.partial/*.download), generic scratch (*.tmp/*.temp), editor swap/backup (*.swp/*.swx/*~), and office/lock files (.~lock.*,~$*). Oncreate(filesystem.rs) a transient name is kept purely local — no empty remote node is minted — and its queued create is parked:next_attempt_atis set to the newPARK_UNTILsentinel soDb::next_due_opnever selects it. Written bytes still ride on that create via the existingattach_blob_to_create, whosenext_attempt_atreset now preserves a park (SQLCASE … >= PARK_UNTIL), so a growing download attaches revision after revision without ever waking the drain. The finalizerename(foo.crdownload → foo) goes through the existingis_local_uidbranch (rewrite_op_targetrenames the queued create); when the old name was transient and the new one is not, it callsDb::set_create_hold(uid, false)and wakes the drain, so the completed file uploads exactly once. Net: no partial and no abandoned temp ever reaches Drive, and nothing forks. Reuses the whole create/attach/rename/drain pipeline — the only new surface is the predicate, thePARK_UNTILsentinel,set_create_hold, aholdarg onqueue_local_node, and the attach-preserves-parkCASE. Tests:syncignorepredicate cases +db::a_parked_transient_create_stays_off_the_drain_until_finalized. No SDK dependency.- Not yet covered by A: apps that write the final name in place with no
temp suffix (rare), and Firefox multi-connection
.partfiles that it may rename per-segment — those still rely on B. An abandoned transient file (a cancelled download left on disk) now stays local-only forever rather than polluting Drive; a later janitor could reclaim its staged blob.
- Not yet covered by A: apps that write the final name in place with no
temp suffix (rare), and Firefox multi-connection
- B — don't self-conflict a single-writer sequential rewrite (done 2026-07-25,
no SDK dep). The safety net for names A does not recognise (an app writing the
final name in place, Firefox per-segment
.partrenames, or any resume the in-memory rebase machinery missed). The daemon now records the server revision id it itself seals per node (Core.own_sealed_revs, written inrefresh_after_uploadwhere the sealed node is already fetched, dropped on trash). When a queued write drains andrevision_changedfires,Core::revision_conflictchecks whether the remote sits at one of our own sealed revisions (pureis_own_self_supersede): if so, no other device touched the file — it is a single-writer stall→resume — so it chains (supersedes) instead of forking a(sync-conflict)copy. Gated onmeta.complete: an incomplete blob's gaps still refer to the stale base, so it keeps the non-destructive conflict-copy path. This is a superset guard over the existingrefresh_after_upload/rebaseline_pendingrebasers, which only fire while a handle/op is live; B catches the windows they miss (fresh open during the earlier upload's flight, a stale inherited baseline). Tests:is_own_self_supersedecases indrain.rs. Residual: the map is in-memory only, so a daemon restart between the partial seal and the resume drops the record and that one write can still fork — narrow, and B69's revision-id keying already covers the mtime-drift half. A persisted own-seal ledger would close it.
Triage of the copies already made is done (2026-07-25): all completed files
were promoted back to their canonical names (metadata-only rename, no re-upload)
and the abandoned .crdownload stubs trashed.
B69 — Spurious (sync-conflict) copies from mtime-keyed revision identity (B16/B25 root fix)
Status: Fixed — offline-tested 2026-07-25, live-validated 2026-07-26 (see below)
Found: 2026-07-24, user reported test_export (sync-conflict 1784898786).xml created after a normal save from an application into the on-demand mount. A full-mount scan turned up ~156 conflict copies; 151 were byte-identical to their live sibling.
Where: crates/pdfs-fuse/src/{drain,filesystem,lib,state,sweep,sync,sync/planner}.rs, crates/pdfs-core/src/cache.rs, crates/pdfs-core/src/control.rs; SDK proton-drive-rs node.rs/client.rs/public_link.rs.
Cause: Two independent false-positive sources, both from (mtime, size) being used as a revision identity (the weakness B25 flagged; the recurrence B16 could not fully close):
- Re-stamped mtime.
revision_conflictcompared the queued write's baseline mtime against the remote's. The server stamps a sealed revision with its own commit time (…780), a few seconds off the optimistic time the client baselined at (…777). Same bytes, same size, drifted mtime → the drain diverted the write into a conflict copy of its own base. - First-sync of pre-existing files.
plan_fileclassified(local, remote, no-baseline)asConflict, so the first reconcile of a folder cut a conflict copy of every file that already existed identically on both sides — the ~151-file storm, all in one ~40s window.
Fix: Make conflict detection key on the server revision id, the true identity (it advances iff a new revision was sealed):
- SDK exposes
active_revision_idandcontent_sha1(plaintext SHA-1 from the revision's decryptedXAttr) onNodeKind::File— both already on the wire / already decrypted, no extra round trips. Baselinegainsrevision_id, captured at write-open (WriteHandle.base_revision_id) and carried across supersede/rebaseline like the mtime.revision_changed(extracted, pure, unit-tested) trusts the id when both sides have it and only falls back to(mtime, size)for an old sidecar. A re-stamped mtime on the same id is no longer a conflict; a same-size edit by another device (new id) still is.- Planner adds
FilePlan::AdoptBaseline: a no-baseline pair of equal size is recorded into the baseline instead of conflicted. Divergent sizes stay a real conflict. - Auto-sweep (
sweep.rs,pdfs-conflict-sweepthread, 5-min cadence): removes conflict copies proven identical to their live sibling (equal size and equalcontent_sha1, trashing the copy remotely) and surfaces divergent/orphaned ones once as a newActivityKind::Conflictentry. Never removes a copy it cannot prove is a duplicate.
Verified: proton-sdk-rs 50/61 lib tests + clippy clean; client cargo fmt/clippy -D warnings/cargo test --workspace --locked all green (258 tests, incl. new revision_changed, planner AdoptBaseline, and conflict_base_name cases). The 151 identical copies were removed manually during triage; the 5 divergent ones are left for the user.
Live validation (2026-07-26): 1.1.0 installed over the production account (6
FUSE mounts of real data), daemon restarted, sweep observed across its warmup pass
and a full 5-minute interval. Pre-run snapshot: exactly 2 (sync-conflict) copies
left on the account, both divergent in size and content_sha1, so the correct
behaviour was to remove nothing. Result: nothing removed — the conflict
manifest was byte-identical before and after, the activity feed gained zero
auto-removed duplicate entries, and the copies were flagged differs from … as
intended. (The 40 trash entries in the feed for that window are
trashed from the mount — a previous FUSE-acceptance run's queued cleanup ops
replaying through the drain, unrelated to the sweep.) The sweep's removal path
therefore remains unexercised against a real account; it is now report-only by
default (B71) precisely so that stays true until there is field evidence.
Shipping note (resolved 2026-07-25): the client's revision-id/sha1 use depends on new SDK surface (proton-sdk/proton-drive-rs 0.2.2). That version is now published to crates.io; the client's workspace deps were bumped "0.2" → "0.2.2", the Cargo.lock re-resolved to the registry crates, and the local [patch.crates-io] shim removed. No local patch remains. (Live validation of the fix against a real account is still pending.)
B68 — Profile backup cannot be written to the device root
Status: Partly fixed — the destination is fixed and verified live
(2026-07-31, see B78); backup health is still invisible in the UI and the
replacement-machine restore has not been re-run
Found: complete managed-live mode matrix against the 1.0.0 release binary
Where: crates/pdfs-fuse/src/profile.rs, profile backup destination
Repro: Every sync-folder registration or mode transition attempted the
profile backup and received NotEnoughPermissions (HTTP 422): Cannot create file at the root of a device. The filesystem matrix itself passed, but profile
backup never became durable remotely.
Impact: The recovery feature appears active but does not preserve the device profile, so a replacement machine still lacks modes, pins, mount settings, and the explicit remote-folder mapping. Repeated retries also add noisy warnings.
Required fix/test: Store the profile below a writable application folder rather than directly at the device root, reuse that folder by stable identity, and make backup health visible. Verify upload, replacement, restart restore, and byte-identical fresh-state recovery on a dedicated account.
B67 — Fresh-state restore can silently select a new device (HIGH-22)
Status: Open — 1.0 recovery blocker
Found: 2026-07-22, 1.0 disaster-recovery audit
Where: device discovery/adoption; docs/RECOVERY.md
Cause: Restore depends on hostname and folder basename. After local-state loss, a mismatch can silently create/select another device rather than attach the intended remote roots, making recovery appear complete while data is absent.
Required fix/test: Require explicit adoption/confirmation on ambiguity and show old/proposed device and root IDs. Test state loss, hostname change, duplicates, renamed folders, multiple devices, and byte-identical restoration.
B66 — Local-only pending data is not protected at shutdown (HIGH-21)
Status: Open — 1.0 data-safety/UX blocker Found: 2026-07-22, 1.0 recovery audit Where: shutdown, CLI/GUI status, logout/unregister workflows
Cause: Undrained writes may exist only on one machine, but there is no flush command with a completion contract, prominent pending byte/count health, or shutdown/logout warning/inhibition.
Required fix/test: Add pdfs sync flush, expose pending operations/bytes
and last errors, and protect destructive lifecycle actions. Test offline stop,
logout, unregister, reboot, retry exhaustion, and drain. Credential removal must
never delete staged bytes.
B65 — Release artifacts lack supply-chain verification (MED-11)
Status: Open — 1.0 release/security blocker Found: 2026-07-22, 1.0 supply-chain audit Where: CI and release workflows
Cause: Releases do not enforce advisory/license policy or publish signed provenance, SBOMs, and verifiable checksums for exact distributed artifacts.
Required fix/test: Add locked audit/license gates, SBOM and provenance, checksums and signatures, documented verification, and an independent clean-job verification/install test.
B64 — Stable publishing lacks data-safety and recovery gates (CRIT-12)
Status: Partly fixed — automated gates and managed-live matrix pass; recovery approval remains open
Found: 2026-07-22, 1.0 release audit
Where: CI and .github/workflows/release.yml
Cause: A tag can publish without required format/lint/tests, offline and dedicated-account live acceptance, migrations, or crash/recovery drills. Fixes still marked unverified can ship as stable.
Required fix/test: Publish immutable RC artifacts only after all workspace checks, full managed mode matrix, migration/power-loss recovery, clean install/ upgrade tests, and explicit sign-off for every critical/high integrity issue.
B63 — Release version identity is inconsistent and unenforced (HIGH-20)
Status: Partly fixed — tag/workspace/PKGBUILD equality is enforced; protocol identity remains Found: 2026-07-22, 1.0 release audit Where: manifests, client identity constants, tags, release workflow
Cause: Workspace/binary versions, protocol identity strings, and tags can
disagree; the workflow trusts any v* tag without checking package metadata or
installed --version output.
Required fix/test: Establish one version source and fail CI on every tag, manifest, package, and binary mismatch. Document any intentionally independent protocol identity.
B62 — Debian artifact omits required service and autostart units (HIGH-19)
Status: Fixed in workflow; clean-install VM verification pending
Found: 2026-07-22, 1.0 packaging audit
Where: .github/workflows/release.yml; packaging/
Cause: The generated package omits proton-drive.service and the tray
autostart entry even though normal login/GUI flows assume that lifecycle.
Required fix/test: Package all units/resources with correct paths, modes, and hooks. In clean supported-distribution VMs test install, login, reboot, daemon/tray/FUSE, upgrade, rollback, and uninstall without deleting user state.
B61 — Rotated single-use credentials can be lost (HIGH-18)
Status: Fixed, not verified live
Found: 2026-07-22, 1.0 authentication audit
Where: crates/pdfs-core/src/auth.rs, refresh callback
Cause: Failure to serialize or persist rotated credentials is only logged (and some setup failures are silent). The daemon continues in memory while the keyring retains an invalid single-use refresh token, so restart loses login.
Required fix/test: Expose unhealthy persistence, retry with bounded backoff, notify the user, and define a safe re-login/shutdown path. Test locked/full/ unavailable keyrings, callback races, repeated rotations, death, and restart.
Fix (2026-10-09, issue #35). A rotation the keyring refuses is kept as unsaved and retried
in the background after 1 s, 5 s, 30 s and then every 5 minutes, until it is written or the blob
has moved on (B201). The refusal is an error in the log naming pdfs login, and the daemon's stop
writes a rotation still unsaved, logging an error if that fails too. The re-login path is B201's:
an old session never writes into a newer login.
Until it is written, the rotation also waits in the kernel keyring (kernelkey::UNSAVED_TOKENS),
so a crash or SIGKILL before a retry lands no longer loses it. The next resume, by the daemon or
pdfs unlock, puts it back in when the blob still holds the refresh token it spent, and writes
it. One that belongs to an older login is dropped. The kernel keyring lasts until the user's last
session ends, so a refused write followed by a crash and a logout still loses the rotation.
While a rotation waits, Response::Status carries tokens_unsaved. The tray says "Sign-in not
saved — unlock your keyring" and needs attention, the app's status says the same, pdfs status
prints a Keyring line and pdfs diagnose warns.
Test: auth::tests::a_rotation_the_keyring_refused_is_written_once_it_takes_it and
auth::tests::an_unsaved_rotation_goes_only_into_the_blob_it_was_refused_for, and the tray's
tests::a_sign_in_the_keyring_refused_needs_attention.
B60 — Config writes are non-atomic and parse errors are overwritten (HIGH-17)
Status: Fixed in code; fault/concurrent-save verification pending
Found: 2026-07-22, 1.0 configuration audit
Where: crates/pdfs-core/src/config.rs, load_config / save_config
Cause: Save truncates in place without sync. Load treats read/parse failure as absence and best-effort overwrites with defaults, silently losing mountpoint, ignore policy, budget, and explicit device adoption.
Required fix/test: Atomically publish a restricted, synced temporary file
and sync its directory. Preserve malformed input and report an actionable error.
Test invalid/truncated JSON, permissions, ENOSPC, crash points, and concurrent
saves.
B59 — Private state permission enforcement fails open (HIGH-16)
Status: Fixed in code with symlink regression; wrong-owner integration verification pending
Found: 2026-07-22, 1.0 local-privacy audit
Where: crates/pdfs-core/src/config.rs, AppDirs::ensure
Cause: chmod/creation failures are warned about or ignored while decrypted content, plaintext metadata, and the authenticated socket may remain accessible to other local users.
Required fix/test: Before loading credentials or serving data, verify every sensitive directory is owned by the effective user, is not a symlink, and is owner-only. Test wrong owner, symlink substitution, read-only filesystems, permissive restored directories, and chmod failure. This tightens B6/B35.
B58 — Control socket has unbounded frames and connections (HIGH-15)
Status: Fixed in code with frame-limit regressions; connection-flood verification pending
Found: 2026-07-22, 1.0 IPC audit
Where: crates/pdfs-fuse/src/control.rs
Cause: The server uses unbounded read_line, a new thread per connection,
and no concurrency or idle limit. A local process can exhaust memory, file
descriptors, and threads, blocking filesystem control and safe shutdown.
Required fix/test: Bound frames, clients, and idle lifetime. Test slow/no- newline clients, oversized/truncated JSON, floods, and recovery. B6/B35 cover socket authority and permissions, not resource bounds.
B57 — Event cursor can advance past unapplied state (HIGH-14)
Status: Partly fixed — cursor now advances per successfully invalidated event; DB/apply fault tests remain
Found: 2026-07-22, 1.0 event-recovery audit
Where: crates/pdfs-fuse/src/background.rs, run_event_sync
Cause: Cache invalidation failures are ignored while the batch cursor still advances. Cursor-persistence failure also leaves the in-memory cursor moving; first-seed persistence failure can later reseed at a newer head. Applying state and committing its cursor have no atomic relationship.
Required fix/test: Apply idempotently and persist derived DB state plus the cursor atomically, or advance only through the last durable event. Retry failed prerequisites. Crash/fault-test every event and cursor boundary, first seed, duplicate replay, malformed events, and database-full conditions.
B56 — Database schema version parsing fails open (HIGH-13)
Status: Partly fixed — future schemas are refused; malformed-version handling remains open
Found: 2026-07-22, 1.0 database audit
Where: crates/pdfs-core/src/db/migrations.rs, Db::migrate
Cause: Missing, malformed, or non-numeric version values become zero, while all future versions are accepted. Corruption can replay migrations, and an older binary can open a newer incompatible database.
Required fix/test: Distinguish a new DB from corrupt metadata; reject malformed/future versions without mutation and validate required schema. Test empty, malformed, negative, future, partial/interrupted migration, and upgrade fixtures from every released schema.
B55 — Incomplete local scans are interpreted as deletions (CRIT-11)
Status: Fixed, with injected iterator/stat failure tests (2026-08-17).
Found: 2026-07-22, 1.0 mirror-sync audit
Where: crates/pdfs-fuse/src/sync.rs, local scan/reconciliation
Cause: Directory-entry and metadata errors omit paths instead of marking the snapshot incomplete. Reconciliation can treat omissions as intentional local deletions and propagate them remotely.
Required fix/test: Carry completeness/errors into planning; an incomplete
subtree must make the pass non-destructive and retain its prior baseline. Test
readdir, stat, permission, transient-I/O, and disappearing-entry failures.
B42 covers the inverse local-delete failure.
Verification (2026-08-17). The scan body was lifted out of Core::walk_local into a free
walk_local_tree, so it can be driven against a real tree with a real unreadable directory in it
and no Core at all. Every failure in it is fatal to the pass by construction — a read_dir that
fails, an entry that cannot be stat-ed, a name that is not UTF-8 — because a skipped path is a
path reconciliation reads as a deletion and propagates to Drive.
Tests: an_unreadable_subdirectory_fails_the_scan_instead_of_shrinking_it (chmod 000 on a
subdirectory holding a file, asserting the error names the subtree),
an_unreadable_root_fails_the_scan, an_absent_root_fails_the_scan (the unmounted-device case),
a_non_utf8_name_fails_the_scan, and a_readable_tree_scans_completely as the complement — the
strictness must not cost a working pass, so a healthy tree still reports every entry and still
skips symlinks by design.
B196 later relaxed this: a path the user may not read, an entry that vanishes mid-scan and a name that is not UTF-8 are left out of the pass instead of failing it, and the pass drops everything under them from classification, so the B55 guarantee holds without the wedge.
Not covered: transient I/O errors, and an entry that disappears between read_dir and stat.
Both are races that would need a filesystem shim to force; the code path they take is the same
symlink_metadata failure the permission test drives.
B54 — Total-wipe guard does not protect a one-entry baseline (CRIT-10)
Status: Fixed, with one-entry, absent-root, unreadable-root and replaced-mountpoint
regressions (2026-08-17). The live mode matrix is still the only thing unexercised.
Found: 2026-07-22, 1.0 mirror-sync audit
Where: crates/pdfs-fuse/src/sync/planner.rs
Cause: The safeguard activates only when the baseline has at least two paths. A one-file local folder that is missing, unreadable, or wrongly mounted can therefore trash its sole remote file.
Required fix/test: Protect every non-empty baseline and require an independently complete scan plus an explicit deliberate-delete signal for a total wipe. Test absent roots, replaced mountpoints, permissions, and one-file folders in every mode combination.
Verification (2026-08-17). The guard reads !baseline.is_empty(), so every non-empty baseline
is protected; guard_local_wipe_blocks_only_a_total_disappearance already carried the one-entry
case. The "independently complete scan" half is B55's fix — an absent, unreadable or partially
readable root now fails the pass before the guard is reached, which those tests cover.
What was missing is the case where nothing fails: a root that is readable but is no longer the
tree it was, an empty directory left where a mount used to be.
a_replaced_mountpoint_trips_the_guard_at_every_baseline_size walks a real empty root, confirms
the scan honestly reports nothing, and asserts the guard fires for baselines of one, two and three
paths — then that it steps aside the moment the tree is back, so a remounted folder resumes
instead of staying wedged.
Deliberate divergence from the required fix: there is no explicit deliberate-delete signal. The guard refuses a total wipe outright rather than offering a way to confirm one. That is stricter than asked for, and it means a user who really did delete everything must remove and re-add the sync folder. Worth revisiting if anyone hits it; not worth building a confirmation path nobody has asked for yet.
Not covered: the live mode matrix (scripts/fuse-acceptance.sh --managed-live).
B53 — Superseding a pending operation is not transactional (CRIT-09)
Status: Fixed, with rollback, restart and full-database verification (2026-08-17).
Found: 2026-07-22, 1.0 SQLite durability audit
Where: crates/pdfs-core/src/db/ops.rs, Db::enqueue_op
Cause: The old durable row is deleted before its replacement is inserted, outside a transaction. A crash, constraint error, full disk, or I/O failure can erase acknowledged upload work and orphan its blob.
Required fix/test: Select and replace in one transaction; retain the old blob until commit and make cleanup recoverable. Inject insert/commit failures and prove that old or new operation—never neither—remains queued. B27 concerns a retained failing row; this issue erases the row.
Verification (2026-08-17). The existing failed_superseding_insert_keeps_the_old_pending_op
proves the row rolls back. Two tests were added for the parts that make the rollback usable.
a_rolled_back_supersede_keeps_its_blob_and_survives_reopen asserts the failed call does not
report the old blob as superseded — that value is the caller's instruction to delete those bytes,
and the row that owns them is still queued — and that the rollback is still there after closing
and reopening a file-backed database, so it is not an artefact of the connection that failed.
a_full_database_rolls_the_supersede_back_rather_than_erasing_it caps max_page_count at the
current size, which is the same SQLITE_FULL an out-of-space filesystem produces, and asserts the
acknowledged upload is still queued and still owns its bytes.
B52 — Staged-write publication is not crash-safe (CRIT-08)
Status: Fixed, with cross-filesystem verification against a real device boundary
(2026-08-17). True power-loss testing is still not covered.
Found: 2026-07-22, 1.0 crash-consistency audit
Where: crates/pdfs-core/src/cache.rs, staged-write publication
Cause: Rename/copy and direct sidecar writes lack a complete sync protocol. The cross-filesystem fallback can remove its source after an unsynced copy, leaving no complete data/metadata pair after a crash.
Required fix/test: Publish synced temporary data and metadata via atomic renames and directory syncs; delete the source only after verification. Inject failure after every write, sync, rename, DB enqueue, and deletion. Related: B50 and B51.
Verification (2026-08-17). stage_write gained a stage_write_at seam — the public entry
point picks a timestamped destination, which is right for production and impossible to arrange a
failure at, the same reason preserve_write_at exists.
staging_across_a_device_boundary_removes_the_source_last drives the copy path across a real
device boundary and asserts the destination and sidecar are both on disk before the source goes.
a_failed_staging_sidecar_keeps_the_cross_device_source blocks the sidecar's temporary name and
asserts the source survives — a staged blob nothing describes is not a substitute for the bytes.
an_unaccompanied_staged_blob_survives_a_restart asserts startup leaves an undescribed staging
blob (and a leftover .json.tmp) alone rather than tidying user data away.
Not covered: real power loss, and injected failure at every one of the listed points — the
sidecar and source-removal boundaries are covered, the intermediate fsync calls are not.
B51 — Scratch sidecar does not establish a durable recovery record (CRIT-07)
Status: Fixed, with crash-boundary and corrupt-sidecar verification (2026-08-17). True
power-loss testing (a machine actually losing power mid-write) is still not covered.
Found: 2026-07-22, 1.0 crash-consistency audit
Where: crates/pdfs-core/src/cache.rs, scratch durability markers
Cause: A temporary sidecar is renamed without syncing the sidecar first or
the directory afterward. Power loss after successful user fsync can preserve
bytes but lose authored-range metadata; recovery may discard the write or treat
a sparse partial write as complete and replace untouched ranges with zeros.
Required fix/test: Sync data, write and sync temporary metadata, rename it, then sync the directory before acknowledging durability. Crash-test every boundary and corrupt/partial sidecars. B20 and B23 do not cover this invariant.
Verification (2026-08-17). a_half_written_scratch_sidecar_is_never_read_as_a_record puts on
disk exactly what a crash inside mark_scratch_durable leaves — a .json.tmp and no published
sidecar — and asserts the next open recovers nothing from it, so an unpublished temporary can
never be replayed as a write. a_corrupt_scratch_sidecar_still_rescues_its_bytes truncates a
published sidecar to a JSON prefix and asserts the blob is still moved to recovery/
unaccompanied: metadata loss must not take the user's bytes with it.
remarking_a_scratch_file_replaces_its_record_atomically covers the repeated-fsync case — each
call leaves one complete parseable record and no debris.
Not covered: real power loss. Nothing short of a crashing VM tests that, and the entry should not be closed as if it did.
B50 — Failed orphan staging deletes the sole scratch copy (CRIT-06)
Status: Fixed, with injected filesystem-failure verification (2026-08-17).
Found: 2026-07-22, 1.0 data-safety audit
Where: crates/pdfs-fuse/src/lib.rs, Core::stage_orphaned_write
Cause: If preserving an orphaned scratch write into staging fails, the failure path removes the original even though no other durable copy exists.
Required fix/test: Never unlink the source until a durable replacement and
metadata are committed. Inject ENOSPC, EIO, EROFS, permission, and
cross-device failures and assert a byte-identical copy remains discoverable.
Related: B23, B27, and B28.
Verification (2026-08-17). Four new tests in pdfs-core/src/cache.rs.
orphan_preservation_keeps_the_scratch_copy_when_staging_is_unwritable makes staging mode 0500
so both the rename and the copy fallback fail, and asserts the scratch file survives
byte-identical, is the path the error names, and carries its recovery marker.
a_retained_orphan_write_is_discoverable_after_a_restart reopens the cache after that failure and
asserts recovered_writes finds the bytes with their metadata — present is not the same as
discoverable, and discoverable is what the entry asked for.
orphan_preservation_across_a_device_boundary_publishes_before_removing and
a_failed_cross_device_preservation_keeps_the_marked_source drive the copy fallback across a real
device boundary (/dev/shm vs /tmp), asserting the ordering — destination and sidecar published
and synced first, source removed last — and that a publication which does not finish leaves the
marked source in place. The cross-device pair skips with a printed note on a machine that has only
one filesystem, rather than passing silently.
Not covered: a genuine ENOSPC/EIO from the block layer; EROFS is approximated by the
unwritable-directory case rather than a read-only mount.
B49 — Failed conflict preservation still overwrites the local file (CRIT-05)
Status: Fixed, with the injected-failure suite the entry asked for (2026-08-17). Only the
end-to-end live conflict remains unexercised.
Found: 2026-07-22, 1.0 data-safety audit
Where: crates/pdfs-fuse/src/sync.rs, download/conflict handling
Cause: If renaming a concurrently changed local file to a conflict copy fails, sync only logs the error and continues publishing the remote download over the local version. Second-resolution conflict names can also collide.
Required fix/test: Abort download publication unless preservation succeeds;
use collision-resistant names. Test existing names, permissions, ENOSPC,
EIO, and cross-filesystem behavior. This is B39's failure path; B25 remains
the separate weak-baseline problem.
Verification (2026-08-17). preserve_conflict_copy was already correct — it copies to a
create_new destination, syncs data and directory, and removes the source last, so every failure
returns Err and apply_one never reaches the download. What was missing was proof, so:
conflict_preservation_failures_all_keep_the_local_file drives an unwritable directory
(EACCES), a source that vanished between planning and preservation (ENOENT), and a destination
name already taken by a directory — and asserts the local file is still there, byte-identical,
after each. a_conflict_copy_is_byte_and_mtime_identical covers the other half: the copy carries
the bytes, mtime and mode, so the next pass uploads the user's version rather than something that
looks freshly written. The existing collision test covers the suffix search.
Not covered: a genuine ENOSPC or EIO from the storage layer, and the end-to-end live conflict.
B48 — Successful unlink can reappear from a stale remote listing
Status: Fixed (unverified)
Found: 2026-07-22, managed acceptance cleanup after all on-demand I/O passed
Where: crates/pdfs-fuse/src/lib.rs, child enumeration and trash paths
Repro: Unlink every child of a directory and immediately remove the
directory. unlink returned success, but a parent invalidation followed by an
eventually consistent remote enumeration returned the just-trashed child again.
rmdir then returned ENOTEMPTY; repeated recursive removal could reproduce it
for the entire acceptance timeout.
Cause: Three stale-data paths combined. A queued cross-directory rename
updated State but did not invalidate the kernel's cached empty destination
listing, so recursive walkers never saw or unlinked the moved child. Removing a node recorded no
authoritative local tombstone, so an eventually consistent Drive listing could
intern it again. Separately, State::has_children treated a resident empty
listing as inconclusive and fell back to obsolete SQLite child rows. Thus
readdir could report empty while rmdir returned ENOTEMPTY forever.
Fix: Every successful online or queued trash records its uid in a
session-shared hidden set. Both persisted and remote child enumeration filter
that set. Secondary on-demand mounts share it, and an explicit restore removes
the uid so restored content can appear normally. has_children now treats any
resident listing—including an empty one—as authoritative and consults SQLite
only when the listing is absent. After replying successfully to a queued rename,
the FUSE handler invalidates both exact dentries and both directory inodes; the
ordering matters because an invalidation sent before the rename reply is
overwritten by the kernel's response processing. Uids are immutable and never
reused, making a session-long tombstone safe.
Test: the acceptance case "regression B48" lists a folder of 12 files, moves one more in,
unlinks them all, and on a mount asks Drive for the listing with pdfs refresh. The listing has to
be empty, and rmdir has to succeed at once.
B47 — FUSE accepts path components longer than NAME_MAX
Status: Fixed and managed-live verified (2026-07-22)
Found: 2026-07-22, managed FUSE acceptance suite
Where: crates/pdfs-fuse/src/filesystem.rs, name-taking callbacks
Repro: open(<mount>/<256 ASCII bytes>, O_CREAT) succeeded. A conventional
Linux filesystem must reject a pathname component longer than 255 bytes with
ENAMETOOLONG.
Cause: Every callback converted OsStr with to_string_lossy and passed it
straight to Drive. This imposed neither Linux's component limit nor Drive's
UTF-8 requirement, and could silently change non-UTF-8 names.
Fix: Route lookup, create, mkdir, unlink, rmdir, and both rename components
through one validator. It enforces the 255-byte limit, rejects reserved/empty
components, rejects invalid UTF-8 with EILSEQ, and preserves valid Unicode.
The managed live suite is the verification gate.
B46 — Combined cross-directory move and rename can fail out of date
Status: Fixed and live-verified (2026-07-22)
Found by: scripts/fuse-acceptance.sh
Where: crates/pdfs-fuse/src/filesystem.rs, ProtonFs::rename
Repro: mv source victim, replacing an existing file, followed immediately
by mv victim dir/moved. Proton accepted the move half of the second operation,
then rejected its rename half with InvalidRequirements (HTTP 422, "out of
date"), surfaced as EIO.
Cause: Moving changes the node's encrypted-name requirements. The following link-details read can briefly observe the pre-move state, so the new name is signed against stale requirements.
Fix history: A bounded retry of InvalidRequirements reduced the window but
did not close it; the managed matrix reproduced the same failure after all four
retries. Combined cross-directory move+rename now enters the durable rename
queue as a desired end state. Its drain re-fetches the node on every attempt,
skips either half that already landed, renames before moving, and resolves name
collisions without losing the source. Simple rename-only and move-only calls
remain synchronous. The managed live suite subsequently passed the replacement
and combined move/rename cases.
Follow-up (2026-09-15, unverified): With SDK 0.6.3 the drain sends the
move and the rename as one move-multiple request carrying the target name,
instead of renaming in the source folder and then moving. That removes the
second call the stale requirements could hit, and the SDK retries a remaining
InvalidRequirements once with the server's current name hash. A destination
name collision now moves under a conflict name in one step. Same-folder renames
are unchanged. The synchronous FUSE rename path (serve_rename) still renames
then moves with its bounded retry and has not been converted. Needs a managed
live run of the combined move/rename cases.
B45 — Truncate after a queued rewrite corrupts or returns EIO
Status: Fixed and live-verified (2026-07-22)
Found by: scripts/fuse-acceptance.sh
Where: crates/pdfs-fuse/src/filesystem.rs, write-open scratch setup;
crates/pdfs-fuse/src/lib.rs, Core::queue_truncate
Repro: Write a file, immediately replace its contents, then run
truncate -s 4 file before the queued replacement drains. Reading it returned
four NUL bytes instead of the first four replacement bytes.
The expanded managed suite found the path-based variant: create and close a
10-byte file, then immediately call truncate(path, 4). The close's complete
revision was still inside the two-second drain debounce. queue_truncate
described the shrink as an incomplete edit over the remote base, noticed the
pending edit, preserved an orphan staging file, and returned EIO.
Cause: A new write handle always treated the server revision as its base. When a newer complete revision was still queued locally, the handle's scratch file began sparse and zero-filled. Shrinking it preserved those scratch zeros, not the prefix of the locally authoritative queued revision.
Fix: A write opened over a complete queued revision seeds its scratch file from that staged blob and marks the copied range authored. Stacking a write over an incomplete queued revision is refused rather than guessed. Path-based truncate now composes the same way: it copies a complete pending blob, applies the new length, marks the result complete, and inherits the last real remote baseline. Live acceptance verified shrink, zero-filled growth, and sparse I/O.
B44 — FUSE background workers have no coordinated shutdown (MED-10)
Status: Fixed and live-verified 2026-08-16 on 1.8.1
Live verification (2026-08-16). systemctl --user restart proton-drive.service on a daemon
with six live mounts, a busy drain and 18.4 GiB staged: stop to Stopped took 110 ms, the journal
shows the six unmounts followed by sync engine stopping and daemon stopping, and systemd never
reached its SIGKILL timeout. The old PID was gone immediately, and the fresh daemon settled at 42
threads — the same count the outgoing one had, so no loop was left behind or spawned twice. A
second restart mid-session behaved identically.
Found: 2026-07-22, daemon lifecycle audit
Where: crates/pdfs-fuse/src/mount.rs, mount lifecycle; drain/index/sync loops
Cause: Long-lived threads and tasks retain Core clones but share no
cancellation token or owned join set. Unmount removes the session/socket without
stopping and joining every worker, so in-process remounts can leak workers and
continue DB/client activity after teardown.
Required fix/test: Add shared cancellation, interruptible waits, owned join handles, and ordered shutdown. Repeated mount/unmount tests should return thread counts to baseline and observe no post-unmount DB work.
Fix (2026-08-16). All four, in crates/pdfs-fuse/src/shutdown.rs plus the loops:
- Shared cancellation.
Shutdown— a flag and a condvar, set once and never cleared — lives onCore, so every clone a worker was given carries it. - Interruptible waits. Every loop that called
thread::sleepin aloop {}with no exit now callsShutdown::sleep, which returnsfalsethe moment teardown starts: the conflict sweep (warmup and interval), the online probe, the local indexer, and both sync poll threads. The drain workers check between ops — never inside one, so an op is always either retired or released — and the sync engine gained aSyncMsg::Stopthat ends its loop and abandons any burst it was settling. - Owned join handles.
run_mountcollects every long-lived thread it starts into oneworkersvec. - Ordered shutdown.
stop_workerssets the flag, wakes the two waits that need waking — the drain's own condvar viawake_drain, and the control listener, which blocks insideacceptand is woken by a connection from us before the socket is unlinked — then joins them all. It runs after the mounts are down, because nothing a worker can still do is harmful at that point, and it also runs on the startup-failure path, which abandons the mount just as completely.
A join that takes a while is expected and correct: a drain worker part-way through an upload finishes it rather than abandoning the user's bytes mid-flight.
Covered by shutdown::tests (a stop cuts an hour-long wait short; an already-stopped signal does
not wait at all). Unverified for the property that motivated the entry — repeated in-process
mount/unmount returning thread counts to baseline — which wants the acceptance suite.
B43 — FUSE directory cookies are not stable (MED-09)
Status: Fixed and live-verified 2026-08-16 on 1.8.1
Live verification (2026-08-16). A probe opened
ProtonDrive/pdfs-live-20260816 (12 files) and called getdents64 with a 160-byte buffer, so the
kernel had to page through the listing; between every page it created a new file in the directory
and unlinked an older one. Three pages, 14 entries returned (12 + . + ..), zero duplicates,
zero pre-existing entries missed — and none of the files created mid-enumeration appeared, which
is the frozen snapshot behaving as specified. The same probe against the pre-fix index-cookie
readdir is what this entry was written about.
Found: 2026-07-22, FUSE/POSIX audit
Where: crates/pdfs-fuse/src/filesystem.rs, ProtonFs::serve_readdir
Cause: readdir rebuilds a live vector for every call and uses its array
index as the continuation cookie. A namespace mutation between pages can shift
indexes and cause entries to be skipped or repeated.
Required fix/test: Implement opendir/releasedir snapshots keyed by the
directory handle, or assign stable per-entry cookies. Test with deliberately
small reply pages and mutation between calls.
Fix (2026-08-16). The first option. opendir is now implemented: it enumerates the folder if
it is cold, freezes the listing, and publishes it under a fresh fh in State::dir_snapshots;
readdir serves from that snapshot; releasedir drops it. The kernel pairs each opendir with
exactly one releasedir, so the snapshot lives exactly as long as the enumeration it belongs to,
and a create or a trash landing between two pages can no longer shift the indexes the continuation
cookie is expressed in.
The fh comes from the same State::next_fh counter write handles use, so a directory handle can
never collide with a file handle. Serving from the snapshot is a pure in-memory walk with no
chance of a remote call, so readdir no longer needs a worker at all; the live path
(serve_readdir) stays as a fallback for a handle that has no snapshot, and both build their
listing through the same build_listing so they cannot disagree.
Enumeration cost is unchanged — it moved from the first readdir to opendir, one call earlier —
and a folder that cannot be enumerated now fails at opendir, where the caller can still act on
the errno.
To verify: the acceptance suite with deliberately small reply pages, creating and trashing
entries between getdents calls, asserting no entry is skipped or repeated.
B42 — Local delete failure is recorded as sync success (HIGH-12)
Status: Fixed (unverified) 2026-08-16.
Found: 2026-07-22, sync error-path audit
Where: crates/pdfs-fuse/src/sync.rs, local file/directory deletion branches
Cause: Some local removal errors are ignored while the baseline is removed and success is logged. The surviving path becomes a new untracked item on the next pass and can resurrect content that was deleted remotely.
Required fix/test: Treat ENOENT as success; for every other removal error,
retain the baseline and increment Outcome.errors.
Fix (2026-08-16). The error half was already in place by the time this was revisited: both
FilePlan::DeleteLocal and the deferred folder removals log the failure, increment
outcome.errors and continue without touching sync_entry_remove, so the baseline survives
and the survivor is not mistaken for a new local file next pass.
The ENOENT half was not, and it failed the opposite way: a path that was already gone was
counted as an error and kept its baseline, so the same removal was retried and re-reported on
every pass forever. Both call sites now go through removed_locally, which reads NotFound as
the state the engine was asking for. Covered by
sync::tests::an_already_absent_path_counts_as_removed. Unverified because no live sync pass has
been driven over an externally-deleted file yet.
Live verification attempted 2026-08-16, blocked. Driving this needs the remote side of a
mirror folder to lose a file, and a mirror folder's remote lives in the device share, which no
mountpoint exposes — pdfs rm only accepts paths under a mountpoint, and the obvious workaround
(switch the folder to on-demand, trash the file through its own mount, switch back) runs into
B86: the on-demand mount serves a listing that predates the mirror engine's uploads, so the
file to be trashed is not visible to rm at all. Verifying B42 live therefore needs either a
second client on the account or B86 fixed first. The ENOENT branch remains covered only by its
unit test.
B41 — Sync folder removal races active reconciliation (HIGH-11)
Status: Fixed (verified by inspection) 2026-08-16 — already correct in the code by the time
this was revisited
Found: 2026-07-22, sync lifecycle audit
Where: crates/pdfs-fuse/src/devices.rs, remove_sync_folder
Cause: Folder removal does not acquire the per-folder sync_lock used by a
reconcile pass. It can unmount, delete configuration/baselines, or trash the
remote root while in-flight tasks continue uploading, downloading, and writing
baseline state.
Required fix/test: Serialize removal with reconciliation, cancel/disable the folder, await active work, then unmount and remove durable state.
Audit (2026-08-16). Core::remove_sync_folder opens by taking self.sync_lock(id) and holds
it across the whole removal, and reconcile_pass takes the same lock for the whole pass — so a
removal requested mid-pass blocks until that pass finishes, which is the serialization this asked
for. The ordering inside the lock is also the one specified: unmount the on-demand session first
(before trashing the remote tree it serves, which would otherwise leave it answering for deleted
nodes), then drop the row.
The one step outside the lock is the remote trash, and that is safe by construction rather than by
accident: the row is already gone, and reconcile_pass re-reads sync_folder_get under the
lock and returns when it finds nothing, so no pass can start against a folder being removed. The
mode re-check on that same re-read is what makes this hold for a switch racing a removal too.
No code change. Recorded as verified-by-inspection: a live removal during an active pass has not been driven.
B40 — Sync upload can stream torn live content (HIGH-10)
Status: Fixed and live-verified 2026-08-16 on 1.8.1
Live verification (2026-08-16). A 128 MiB b40.bin was dropped into folder 6 and a pass
forced; six seconds into the upload, sixteen bytes were overwritten at offset 100,000,000. The
pass logged
WARN pdfs_fuse::sync: sync: file changed while uploading; the remote revision may be torn and
will be replaced on the next pass rel="b40.bin"
streamed=LocalSig { mtime: 1786916599, mtime_ns: Some(1786916599336211782), size: 134217728 }
now=Some(LocalSig { mtime: 1786916605, mtime_ns: Some(1786916605686365985), size: 134217728 })
and — this is the part that matters — recorded the baseline against the streamed signature
(…599), not the file on disk. The very next pass therefore read the file as locally changed and
uploaded b40.bin new version. Round-tripping the folder through on-demand and back later
downloaded the remote copy at md5 cfb45003003fb2ad50c2c365f8313849 — the post-mutation local
content, exactly. The torn revision healed itself without anyone asking.
Found: 2026-07-22, sync concurrency audit
Where: crates/pdfs-fuse/src/sync.rs, upload apply and baseline update
Cause: Planning stats a path, but upload later opens and streams the live file. A concurrent writer can change or truncate it during transfer, producing a torn remote revision. Baseline settlement may then stat still newer local bytes and falsely declare them synchronized.
Required fix/test: Upload an immutable staged snapshot, or validate an open descriptor before and after streaming and refuse baseline settlement on change.
Fix (2026-08-16). The second option, and the correction is structural rather than a check.
record_file_baseline no longer stats the path when it is done; it is given the identity the
transferred content corresponds to, and both upload paths pass the LocalSig taken from the
descriptor they streamed. So a file that changed mid-stream records a baseline describing the
bytes that went up, not the bytes on disk now — which makes the very next pass classify the path
as locally changed and upload a whole, clean revision over the torn one. The old behaviour stated
the opposite: it declared the torn revision synchronized and left it as the file's content on
Drive.
warn_if_torn compares before and after purely so the journal explains the extra revision; the
correction does not depend on it. Downloads are unaffected — there the file on disk is the
transferred content, published by a rename and stamped, so its stat is the identity.
Note the pre-existing mitigation this backs up: the /proc/*/fd scan already defers any file held
open for writing (FilePlan::Deferred), so this covers the residual — a writer that opens, writes
and closes entirely inside one upload.
To verify: upload a large file from a mirror folder while appending to it, then check that the following pass uploads a new revision and the final remote content matches the local file.
B39 — Sync download can overwrite a concurrent local edit (HIGH-09)
Status: Fixed and live-verified 2026-08-16 on 1.8.1
Live verification (2026-08-16). Folder 6 was switched to on-demand (evicting its local
copies) and then back to mirror, which queues a download of the 128 MiB b40.bin. Three seconds
into that download, RACING-LOCAL-EDIT-B39 was written to the destination path. The pass logged
WARN pdfs_fuse::sync: sync: local file changed while downloading; keeping it as a conflict copy
path=/home/narl/pdfs-live-mirror/b40.bin expected=None
actual=Some(LocalSig { mtime: 1786916720, mtime_ns: Some(1786916720414151922), size: 21 })
and finished with the remote copy at b40.bin and the racing 21 bytes intact at
b40 (sync-conflict 1786916794).bin — which the following pass then uploaded as a file of its
own. Before the fix the rename would have destroyed those bytes with no trace.
Found: 2026-07-22, sync concurrency audit
Where: crates/pdfs-fuse/src/sync.rs, download classification and apply path
Cause: Reconciliation classifies a local file from an earlier scan, downloads to a temporary file, then renames over the destination without revalidating that the local inode/content stayed unchanged. A writer racing the download can have its completed edit silently replaced.
Required fix/test: Capture a local identity/signature during planning and revalidate immediately before rename. Preserve a conflict copy on mismatch. A deterministic test should pause the download, edit the target, resume it, and verify that neither version is lost.
Fix (2026-08-16). Pending::Download now carries local: Option<LocalSig> — what the
planning walk actually saw at that path, or None for nothing there — and download_file calls
keep_racing_local_edit in the last moment before the rename that publishes the download. A
destination that no longer matches the plan is moved aside by preserve_conflict_copy first, so
the racing bytes survive and re-upload as a new file on the next pass: the same resolution a
both-sides-changed conflict already gets. A path that is empty when something was expected is
not a conflict — the download is about to recreate it and there is nothing to preserve.
Pending::Conflict passes None deliberately: it has just moved the local copy aside itself, so
the state its download expects to find is "nothing", and anything that appears while it runs is a
new racing write that gets its own copy.
Covered by sync::tests::a_download_keeps_a_local_edit_it_did_not_plan_for (unchanged destination
overwritten silently; changed destination preserved). Unverified against a real timing race —
the deterministic pause-mid-download test the entry asks for is not in place; the unit test drives
the guard directly instead.
B38 — Sync task panic could authorize destructive mode switch (HIGH-08)
Status: Fixed in code (2026-07-22)
Found: 2026-07-22, sync concurrency audit
Where: crates/pdfs-fuse/src/sync.rs, Core::flush_batch
Cause: A JoinSet task panic was logged but not added to Outcome.errors.
The pass could therefore be recorded as idle/successful even though an upload or
download never ran, allowing a pending on-demand switch to evict local data.
Fix: Count every join failure as a reconciliation error, which prevents the pass from being treated as successfully settled.
B37 — Failed remote trash removed the local dentry (CRIT-04)
Status: Fixed in code (unverified on a live mount, 2026-07-22)
Found: 2026-07-22, deep FUSE/POSIX audit
Where: crates/pdfs-fuse/src/filesystem.rs, ProtonFs::trash_child
Cause: trash_child called forget_or_unlink before the online
trash_nodes request. If Drive rejected or failed the request, FUSE returned
EIO but the path had already disappeared from local state.
Fix: Queue or complete the remote mutation first and remove the local dentry only after that step succeeds. Offline deletion still becomes immediately visible, but only after its durable queue row exists.
B36 — Unicode normalization discrepancy (NFC vs NFD) (LOW-01)
Status: Open (verified 2026-08-16, deliberately not fixed)
Found: 2026-07-21, multi-agent sync audit (audit_bugs.md LOW-01)
Where: crates/pdfs-fuse/src/lib.rs, crates/pdfs-fuse/src/sync.rs
Cause: macOS/HFS+ NFD UTF-8 path inputs differ from Linux NFC UTF-8 paths, causing duplicate folder creation or lookup misses for accented filenames.
Verified (2026-08-16), not fixed. The claim is accurate: names are compared as bytes
throughout — the mirror engine's rel paths, by_uid/children interning, and the planner's
classification union — so café written by a Proton client on macOS or iOS (NFD) and café
written here (NFC) are two different paths, and a mirror folder holding both would sync both.
Not fixed here on purpose. A normalizing comparison changes what "the same file" means across the whole daemon, and it changes it in the unsafe direction: two names that are genuinely distinct on the remote would start colliding locally, and a collision in the sync engine is a deletion or an overwrite, not a warning. Doing it properly means a normalization dependency, a decision about which form is canonical on the wire, and a migration for baselines keyed on the other form — its own change with its own acceptance case, not a line in a sweep. Left LOW because it needs cross-platform authorship of the same accented name to bite at all.
B35 — IPC Unix socket creation permission race (MED-08)
Status: Fixed (verified 2026-07-20 - see B6)
Found: 2026-07-21, multi-agent security audit (audit_bugs.md MED-08)
Where: crates/pdfs-fuse/src/control.rs, crates/pdfs-core/src/config.rs
Cause: Binding the domain socket before setting chmod(0600) created a race window where local users could connect before permissions were enforced.
Fix: Fixed via B6 remediation: AppDirs::ensure sets 0700 permissions on config, state, and cache directories, and config::restrict_socket applies 0600 immediately after binding control and tray sockets.
B34 — Writable POSIX exposure of read-only shared folders (MED-07)
Status: Fixed in code by P2/P3 (2026-07-29); P3 now exposes the shared tree, while credentialed
second-account and live FUSE acceptance remain pending P7
Found: 2026-07-21, multi-agent permissions audit (audit_bugs.md MED-07)
Where: crates/pdfs-fuse/src/lib.rs, crates/pdfs-fuse/src/sharing.rs
Cause: Shared "Viewer" (read-only) folders are exposed via POSIX with write permissions (0755). Local writes succeed initially on the mount but fail perpetually during background online drain with 403 Forbidden errors.
Fix: P2 persists share authority in schema V17 share_access, inherits it through resident
entries, exposes read-only POSIX modes, and gates mutations at handler, queue, and drain boundaries.
Downgrades fail closed and are not acknowledged past failed durable state updates.
Verification: The local workspace passes formatting, locked checks, clippy with warnings denied, and all 347 workspace tests, including P3 virtual-tree materialization and access inheritance plus mutation gating, queued-operation authorization, and downgrade failure paths. Credentialed second-account and live FUSE acceptance have not been run and remain required in P7.
B33 — Lack of partial transfer resumption & head-of-line queue blocking (MED-06)
Status: Fixed (verified 2026-07-22)
Found: 2026-07-21, multi-agent drain engine audit (audit_bugs.md MED-06)
Where: crates/pdfs-fuse/src/drain.rs:L73
Cause: The attempted remediation classified errors by substring and deleted the pending row after five failures or strings containing 402/403/413. There was no quarantine table or export path: the staged blob became unreachable and the in-memory pending view could remain stale. Five transient failures are not proof that accepted user data is disposable.
Fix: Never delete an accepted write on a retry path. Every failure remains a
durable pending operation and receives bounded exponential backoff; the indexed
next_due_op query skips it until due, so unrelated work is not blocked.
Verified: cargo clippy -p pdfs-fuse --all-targets -- -D warnings and the
pdfs-fuse unit suite pass. A durable user-visible quarantine feature may be
added later, but it must retain the row and blob.
B32 — Race condition in concurrent handle release and open (MED-05)
Status: Fixed and live-verified 2026-08-16 on 1.8.1
Live verification (2026-08-16). Three 64 MiB random files were written over the same mount
path (ProtonDrive/pdfs-live-20260816/b32.bin) back to back, each cp closing before the next
began, while the drain was still working. The queue coalesced them into a single op (29 queued =
28 known-stuck + 1). After the drain retired it, pdfs refresh dropped the cached listing and a
full re-read from Drive returned 4ad41482f5199a401e425381d7621626 — the md5 of the third
write, byte for byte. No open re-used a stale staged base, and the retry warning never had to
fire.
Found: 2026-07-21, multi-agent concurrency audit (audit_bugs.md MED-05)
Where: crates/pdfs-fuse/src/lib.rs:L4042, L4574
Cause: Opening a file for writing while a previous handle release is draining creates overlapping scratch write handles and out-of-order remote revision uploads.
Fix (2026-08-16). The concrete hole was in serve_open: it sampled the queued revision
(Core::pending), then copied that blob into a fresh scratch file, then took the State lock to
install the handle. A release publishing a newer revision anywhere inside that gap — which spans
a copy the size of the whole file — left the new handle carrying the older base, and its own
release then published content that silently discarded the revision closed a moment earlier.
The install is now guarded: the pending blob path (which changes on every publication, so it
identifies the revision) is re-read immediately before the State lock, and if it moved, the
scratch file is discarded and the open goes around again — up to OPEN_BASE_ATTEMPTS (3), after
which it fails with EIO rather than spinning a FUSE worker. Only a handle the open is about to
create is checked; if one already exists, that handle is authoritative and the scratch is
discarded regardless.
The re-read is deliberately not taken under the State lock: no site in the daemon holds
pending and state at once (the invariant is stated on stamp_pending_sizes) and this is not
the place to become the first. What that costs is a residual window of a few instructions instead
of a lock-tight comparison; what it removes is the multi-gigabyte window, which is the one that
actually loses data.
To verify: the acceptance suite's concurrency case, with a large file, closing and reopening it for write from two processes while the drain is loaded.
B31 — Unimplemented special files & attribute modifications (MED-04)
Status: Not a bug — by design, verified 2026-08-16
Found: 2026-07-21, multi-agent POSIX audit (audit_bugs.md MED-04)
Where: crates/pdfs-fuse/src/lib.rs
Cause: Symlinks (symlink/readlink), hardlinks (link), FIFOs/sockets (mknod), and chmod/chown attribute modifications return ENOSYS or are silently ignored.
Verdict (2026-08-16). Each of these is the correct answer, not a gap, and each errno is now a deliberate choice with a comment saying why:
symlink→ EPERM. The operation is meaningful; Drive simply cannot represent it. Callers treatEPERMas a definite "no" instead of retrying against an older interface.link→ EPERM. A node has exactly one parent on Drive, so there are no hard links. git falls back torename()onEPERM, which this filesystem supports.mknod→ ENOSYS. Devices, FIFOs and sockets have no representation at all. Regular files arrive throughcreate, never here.chmod/chown/utimes→ accepted and ignored, so tools that chmod or restamp after writing succeed rather than failing on a filesystem with no per-file mode to set.
Implementing any of them would mean inventing a local-only representation that no other Proton client can read — a worse outcome than a clean refusal. The errno contract is asserted by the acceptance suite's clean-refusal case so it cannot drift, and B73 removed the per-probe WARN that made these look like incidents in the journal.
B30 — SQLite write lock & mutex contention under heavy drain (MED-03)
Status: Fixed (unverified under load) 2026-08-16 — closed by the work filed under the
performance audit rather than by a change made for this entry.
Found: 2026-07-21, multi-agent database audit (audit_bugs.md MED-03)
Where: crates/pdfs-core/src/db/ops.rs, crates/pdfs-core/src/db/mod.rs
Cause: Long-running upload drain transactions hold SQLite write locks while worker threads wait for state.lock(), causing periodic FUSE response latency spikes during heavy sync operations.
Audit (2026-08-16). Each of the three mechanisms behind that sentence is gone:
- No transaction spans a network call. Every
Dbmethod opens its transaction, commits and returns; nothing holds one open across calls (the invariant is stated onDb::read). A drain op's upload happens between database calls, not inside one. - Reads no longer queue behind the writer. The read-only connection pool (
Db::read) givesSELECT-only methods their own connections, so a FUSE metadata call does not wait on whatever is committing. Asserted bydb::tests::a_read_does_not_wait_for_the_write_connection, which holds the write connection for the whole measurement. - The queue is no longer read whole to pick one op.
Db::claim_next_due_opselects and claims a single row in the database, which is what made a long queue quadratic to drain while holding the shared connection.
The state.lock() half is covered from the other side: B75 moved network work off fuser's
dispatch thread, and B22's audit found no block_on under a State guard.
Unverified because "periodic latency spikes" is a load property, not a code property: it is closed by construction and wants the concurrency acceptance case under a loaded drain to confirm it in the field.
B29 — Missing FUSE forget method / memory leak (MED-02)
Status: Fixed (verified 2026-07-22)
Found: 2026-07-21, multi-agent FUSE state audit (audit_bugs.md MED-02)
Where: crates/pdfs-fuse/src/lib.rs:L3983, crates/pdfs-fuse/src/state.rs
Cause: Kernel forget messages sent by FUSE were ignored, causing state.entries and state.by_uid maps to grow monotonically in RAM over long daemon uptimes.
Fix: Added lookup_count to Entry and implemented forget_lookup(ino, nlookup) in state.rs. ProtonFs::forget now invokes st.forget_lookup(ino.0, nlookup) to prune unreferenced inodes from memory.
Verified: cargo clippy --workspace --all-targets and cargo test passing with zero warnings.
B28 — Unhandled disk exhaustion (ENOSPC) in cache & scratch store (MED-01)
Status: Fixed in code — previous remediation was unsafe (2026-07-22)
Found: 2026-07-21, multi-agent storage audit (audit_bugs.md MED-01)
Where: crates/pdfs-core/src/cache.rs, emergency_evict
Cause: Running out of local disk space caused unhandled EIO errors without triggering emergency eviction of unpinned cache blobs.
Fix: Added emergency_evict() to ContentCache which queries cache_eviction_candidates and evicts LRU unpinned blobs immediately to reclaim disk space when ENOSPC occurs.
Verified: cargo test -p pdfs-core suite passing cleanly.
B27 — Orphaned write revision bricking offline sync queue (HIGH-07)
Status: Fixed in code (unverified create→unlink live sequence, 2026-07-22)
Found: 2026-07-21, multi-agent drain engine audit (audit_bugs.md HIGH-07)
Where: crates/pdfs-fuse/src/lib.rs:L1811, crates/pdfs-fuse/src/drain.rs
Cause: Unlinking a newly created offline file discards its Create pending op, but closing an open write handle on the file queued a Write revision for local~....
Fix: In addition to the queue_revision guard, create participates in
inode open-lifetime accounting and release explicitly suppresses revision
queueing when it retires the last handle of an unlinked inode.
Verified: Code inspection in lib.rs:1811 and workspace tests passing.
B26 — Broken open-unlinked file semantics (HIGH-06)
Status: Fixed (verified 2026-07-22)
Found: 2026-07-21, multi-agent FUSE audit (audit_bugs.md HIGH-06)
Where: crates/pdfs-fuse/src/lib.rs, crates/pdfs-fuse/src/state.rs
Cause: Unlinking an open file called trash_child, which invalidated the node and dropped it from state immediately. Subsequent read calls on open file handles failed with ENOENT / EIO.
Fix:
- Added
open_countandunlinkedfields toEntryinstate.rs. openincrementsopen_count,releasedecrementsopen_count.forget_or_unlinkmarksentry.unlinked = trueand removesinofromst.childrenof the parent (so the file disappears from directory listings immediately) while preserving the inode inst.entries.readcalculates optimisticfsizefalling back tocore.pendinglength for unlinked/pending files.- Node removal, op cleanup, and cache eviction are safely deferred until
releasebringsopen_countto 0. createnow increments the returned inode'sopen_count, writable-open scratch allocation failure rolls the count back, and the last release of an unlinked writer discards its scratch bytes instead of queueing a resurrection.
Verified: Live mount verification: python script opened file descriptor, unlinked file via rm, verified file disappeared from directory listing, and read full contents from open fd cleanly on ~/testmount and ~/testmount-2.
B25 — Weak conflict baseline detection (mtime, size) (HIGH-05)
Status: Fixed and live-verified 2026-08-16 on 1.8.1 — the FUSE-write half was closed by B69; the mirror half is closed here.
Live verification (2026-08-16, ~/pdfs-live-mirror, folder 6). b25.txt was written with
mtime 12:00:10.100000000, synced (baseline local_mtime=1786874410,
local_mtime_ns=1786874410100000000), then rewritten with sixteen different bytes and stamped
12:00:10.900000000 — identical whole second, identical length. The next pass classified it as
locally changed and uploaded b25.txt new version; the baseline advanced to
…410900000000, and reading the file back through an on-demand mount of the same folder returned
BBBBBBBBBBBBBBBB. Under the old whole-second comparison this edit was invisible.
Found: 2026-07-21, multi-agent sync engine audit (audit_bugs.md HIGH-05)
Where: crates/pdfs-core/src/cache.rs:L157, Baseline
Cause: Conflict resolution baseline detection relies strictly on (mtime, size) tuple comparisons without content hashing or revision IDs. Concurrent edits landing within the same second that produce identical file sizes match baselines, causing offline edits to overwrite remote changes silently without creating (sync-conflict <ts>) copies.
Fix (2026-08-16), part one — the FUSE write path. Closed by B69: Baseline gained
revision_id, and revision_changed keys on the server revision id — the true identity, which
advances if and only if a new revision was sealed — falling back to (mtime, size) only for a
sidecar written before the field existed.
Fix (2026-08-16), part two — the mirror sync path. B69 did not reach it: a mirror folder's
baseline is sync_entry, and its local side was still whole seconds. An edit made in the same
second as the last sync that left the file the same length — a fixed-size record rewritten in
place, an image re-exported, a database page updated — read as unchanged and was never
uploaded, leaving the remote silently behind until something else about the file moved.
sync_entry gains local_mtime_ns (schema 28), and the comparison moved into
LocalSig::same_content, which uses nanoseconds when both sides have them and whole seconds
otherwise. The column is nullable rather than a converted local_mtime * 1e9 on purpose: the
sub-second part of an existing baseline is genuinely unknown, and inventing zero for it would make
every already-synced file compare as changed and re-upload an entire mirror on first start. NULL
means "compare seconds", which is exactly the behaviour that row was written under; each row gains
a real value the next time its path syncs.
Covered by db::tests::migration_v28_adds_sub_second_local_times_without_disturbing_a_v27_baseline
and sync::planner::file_plan_tests::a_same_second_same_size_edit_is_still_a_local_change.
Still open (deliberately): the remote side of a mirror baseline is still (mtime, size)
(remote_rev/remote_hash). Re-keying it onto active_revision_id the way B69 did for the FUSE
path is the same idea and the same size of change, but the stored column already holds
mtime-shaped strings, and telling an old row from a new one would need a heuristic on the value's
shape. That deserves its own column and its own pass rather than a guess.
To verify: on a mirror folder, printf %s aaaa > f && sync && printf %s bbbb > f inside one
second, then a reconcile pass — expect an upload, and the remote content to be bbbb.
B24 — Directory hierarchy loop in rename (HIGH-04)
Status: Fixed (verified 2026-07-21)
Found: 2026-07-21, multi-agent FUSE audit (audit_bugs.md HIGH-04)
Where: crates/pdfs-fuse/src/filesystem.rs, ProtonFs::rename, crates/pdfs-fuse/src/state.rs:L237
Cause: Moving a directory into one of its own subdirectories created an infinite recursive loop in inode state memory and DB tree structure.
Fix: Added State::is_ancestor_of helper in state.rs that walks up parent chains. rename checks st.is_ancestor_of(ino, newparent) when moving directories and returns Errno::EINVAL if a cycle would be formed.
Verified:
- Unit tests:
test_is_ancestor_of_hierarchyinstate.rspassing. - Live mount verification:
mv parent parent/childreturnedEINVAL("Invalid argument") on both mounts.
B23 — Un-fsynced scratch wipe on daemon restart (HIGH-03)
Status: Fixed (verified 2026-07-22)
Found: 2026-07-21, multi-agent recovery audit (audit_bugs.md HIGH-03)
Where: crates/pdfs-core/src/cache.rs:L741, ContentCache::rescue_scratch
Cause: On daemon startup, rescue_scratch inspected the scratch directory and only preserved files with valid .json sidecars created by fsync. Active writes closed without an explicit fsync lost their sidecars and were deleted during startup directory cleanup.
Fix: Updated rescue_scratch in cache.rs to generate synthetic StagedWrite sidecars for any scratch file with valid data (len > 0) when a sidecar exists but was partially corrupted, preserving offline writes across unclean shutdowns.
Verified: Unit tests fsynced_scratch_survives_reopen_and_unmarked_scratch_does_not and cleared_durability_marker_stops_recovery in cache.rs passing 100%.
B22 — Synchronous block_on inside mutex guard (HIGH-02)
Status: Fixed (unverified under load) 2026-08-16 — the systematic audit this was waiting on
has been done and found nothing left.
Found: 2026-07-21, multi-agent concurrency audit (audit_bugs.md HIGH-02)
Where: crates/pdfs-fuse/src/lib.rs, sync.rs, devices.rs
Cause: Invoking rt.block_on(...) while holding state.lock() deadlocks worker threads and the main FUSE loop under high concurrency. Most major handlers have been refactored to drop state.lock() before invoking async runtime methods, but a full systematic audit across all background tasks and FUSE callbacks is required to eliminate all instances.
Audit (2026-08-16). All 131 block_on sites in pdfs-fuse were scanned mechanically: for
each, every guard binding (self.state() or any .lock()) still lexically live at that point,
accounting for explicit drop(..) and scope exits. Two hits, both false positives —
devices.rs:570 binds taken, the Option<Mount> taken out of the map (the guard is a
temporary, and _guard is explicitly dropped above), and lib.rs:1917/1927 binds resident,
a bool read out of a temporary guard. pdfs-core contains no block_on at all. The reverse
direction — a guard held across an .await — is covered by clippy::await_holding_lock, which
is on in the CI gate.
The scan is lexical, so it cannot see a guard passed into a callee; it is evidence, not proof, which is why this is unverified. What would settle it is the concurrency half of the acceptance suite run against a heavily loaded drain.
B21 — Untracked hole punching in fallocate (HIGH-01)
Status: Fixed (verified 2026-07-22)
Found: 2026-07-21, multi-agent FUSE audit (audit_bugs.md HIGH-01)
Where: crates/pdfs-fuse/src/filesystem.rs, ProtonFs::fallocate
Cause: FALLOC_FL_PUNCH_HOLE zeroed scratch blocks but removed their
WriteHandle::written authored status. Revision assembly therefore classified
the punched range as an untouched gap and refilled it from the remote baseline,
silently undoing the hole at commit.
Fix: Mark the punched, in-file range as authored after successful local
fallocate, so its sparse zero bytes are retained by revision assembly.
Verified:
- Live mount verification:
libc.fallocate(fd, 0x03, 256, 512)zeroed middle range while preserving edge authored bytes on both~/testmountand~/testmount-2.
B20 — Un-fsynced temp file atomic rename in content cache (CRIT-03)
Status: Fixed (verified 2026-07-21)
Found: 2026-07-21, multi-agent storage audit (audit_bugs.md CRIT-03)
Where: crates/pdfs-core/src/cache.rs, ContentCache::store
Cause: Cache store renamed temporary files into the cache directory without explicit fsync. Power loss or ungraceful shutdown could leave 0-byte or corrupted files indexed as valid cache hits.
Fix: Added f.sync_all()? calls on open file handles prior to executing std::fs::rename.
Verified: cargo test -p pdfs-core suite passing cleanly.
B19 — Remote folder trashing on mode switch failure (CRIT-02)
Status: Fixed; all four managed-live mode pairs verified 2026-07-22,
fault-injected mount failure still pending
Found: 2026-07-21, multi-agent sync engine audit (audit_bugs.md CRIT-02)
Where: crates/pdfs-fuse/src/devices.rs, apply_sync_folder_mode
Cause: The first fix mounted FUSE before calling evict_dir_contents. That
made the cleanup walk the newly mounted remote namespace rather than the hidden
local mirror, issuing FUSE unlink/rmdir operations which could trash the
entire remote folder. The local files were not reclaimed at all.
Fix: Persist ondemand first so reconciliation cannot interpret cleanup as
user deletion, evict the underlying local mirror while holding the sync-folder
lock, and only then mount FUSE. A mount failure leaves an inert ondemand row;
switching back to mirror clears the baseline and restores the local copy.
Verified: cargo clippy -p pdfs-fuse --all-targets -- -D warnings, the
pdfs-fuse unit suite, and the complete managed-live matrix passed against the
1.0.0 release binary. A fault-injected mount-failure test is still needed.
B18 — Subtree erasure via POSIX unlink / rmdir (CRIT-01)
Status: Fixed (verified 2026-07-22)
Found: 2026-07-21, multi-agent filesystem safety audit (audit_bugs.md CRIT-01)
Where: crates/pdfs-fuse/src/filesystem.rs, ProtonFs::unlink, ProtonFs::rmdir
Cause: unlink on directories and rmdir on non-empty directories bypassed POSIX checks and called trash_child directly, silently deleting remote subtrees.
Fix: Added type and non-emptiness checks in lib.rs:
unlink: ReturnsErrno::EISDIRif target is a folder.rmdir: ReturnsErrno::ENOTDIRif target is a file, enumerates a cold directory before deciding whether it is empty, and returnsErrno::ENOTEMPTYwhen it contains children. An absent cache entry is unknown, not empty; without the enumeration Proton's recursive trash operation could erase unseen children.
Verified:
- Unit tests:
test_posix_unlink_and_rmdir_checksinlib.rspassing cleanly. - Live mount verification on
~/testmountand~/testmount-2:rm dirreturnedEISDIR,rmdir filereturnedENOTDIR,rmdir non_empty_dirreturnedENOTEMPTY,rmdir empty_dirsucceeded.
B17 — FUSE mount lock contention, missing next_attempt_at in DB enqueue, and ioctl ENOTTY fix
Status: Fixed (verified 2026-07-21)
Found: 2026-07-21, deep code audit of fix/ioctl-handler branch.
Where: crates/pdfs-fuse/src/lib.rs, crates/pdfs-core/src/db/ops.rs
Cause:
Db::enqueue_opSQLINSERT INTO pending_opstatement omittednext_attempt_atfrom the column list, causing SQLite to setnext_attempt_at = 0and rendering the 2-second revision debounce (DRAIN_REVISION_DEBOUNCE) ineffective.ioctlreturnedENOSYSfor non-terminal commands, signaling to kernel FUSE that ioctl calls were unsupported across the mount.fallocateexecutedlibc::fallocatewhile holdingself.core.state.lock(), blocking all concurrent FUSE operations across the mount during disk block preallocation.open/createinvokedcreate_scratch()prior to checkingst.active_writes, leaking scratch files on disk during concurrent opens.
Fix:
- Updated
Db::enqueue_opinops.rsto includenext_attempt_atin theINSERT INTO pending_opstatement and parameters. - Changed
ioctlhandler inlib.rsto returnErrno::ENOTTYfor all unhandled ioctls, conforming to POSIX standards. - Moved
libc::fallocateoutside thestatelock by cloning theArc<File>reference first, and mapped system OS errors tofuser::Errno. - Refactored
openandcreateto checkst.active_writesunder lock before invokingcreate_scratch(), preventing redundant scratch file creation.
Verified: Workspace unit tests (200/200) and clippy passing cleanly.
B16 — aria2c preallocation / rapid sequential writes trigger false self-conflict copies
Status: Fixed (verified 2026-07-21)
Found: 2026-07-21, user reported sync conflict files created during torrent downloads with aria2c onto the FUSE mount ([DB]Oshi no Ko 3rd Season_-_05... (sync-conflict 1784644482).mkv).
Where: crates/pdfs-fuse/src/lib.rs, crates/pdfs-fuse/src/drain.rs, crates/pdfs-core/src/db/ops.rs
Cause: Tools like aria2c preallocate target file sizes using ftruncate/setattr and then write actual content across multiple file opens or handle releases.
- First close (e.g. after preallocation): queued an
OP_REVISIONwith baseline set to the server revision at open time. - If this initial op drained immediately,
refresh_after_uploadupdated the remote node's modification time on the server. - Second write / close (with actual content): if the write handle was opened before the first op drained or if the op drained before the second write released,
remote_baselinebuilt abased_onreferencing the old server revision. - When the second op drained,
revision_conflictcomparedbased_onagainst the new remote revision mtime, detected a mismatch, and treated it as a remote edit by another device — uploading the local data as a(sync-conflict <ts>)duplicate file.
Fix: A two-part defense:
- Revision Debounce (
DRAIN_REVISION_DEBOUNCE = 2s):enqueue_staged_writesetsnext_attempt_at = now + 2sforOP_REVISIONops instead of0. Rapid follow-up writes supersede the queued op in staging before it ever reaches the network. AddedDb::earliest_due_at()andwait_for_drain_work_or_due()so the drain loop sleeps precisely until debounced ops become due. - Open-Handle Rebaselining:
refresh_after_uploadnow updatesbase_mtimeandbase_sizeon any openWriteHandletargeting the same node UID when a revision seals, complementingrebaseline_pending.
Verified: Clean build and 200/200 workspace tests passing.
Residual (2026-07-24): the two-part defense narrowed the window but did not close it — the baseline was still keyed on (mtime, size), which drifts when the server re-stamps the same revision. Reproduced again on a plain application save (test_export.xml): base mtime 1784897777 vs sealed 1784897780, identical size, no other device involved. The root cause is that mtime is not a revision identity. Fixed properly by B69.
B15 — An empty-but-listed folder, and duplicated folder uids
Status: Open (side finding, unattributed). Originally filed as "a sync-folder listing came back empty"; that part was investigated and retracted — see below. Found: 2026-07-20, while confirming B13.
ls ~/Music/Buunshin/not everythin is your fault (wav)/ returns total 0 — no
entries — even though the rename failures in B13 prove files existed under those
names minutes earlier (Proton rejected the rename because the name was taken).
The same folder read through the primary mount
(~/ProtonDrive/Music/Buunshin/…) lists all seven .wav files correctly.
Checked (2026-07-20) — the "stale empty listing" reading was wrong. The
folder not everythin is your fault (wav) that read empty has listed = 0
in nodes, so its ls was a real network enumeration, not a cached empty one.
The files genuinely are not there: every rename in B13 failed and rsync
discarded its temp files. B15 as originally filed does not exist.
What the query did turn up, and is worth keeping:
- A real B1-shaped row.
not everything is your fault(no(wav)) readslisted = 1, trashed = 0with zero child rows — the exact empty-but-listed state B1 was about, still present in the live DB, reached by some route the B1 fix does not cover. This one is worth chasing. - Duplicated folders.
MusicandBuunshineach exist under two distinct uids, and the(wav)folder name appears twice as well — one copy populated (7 children), one empty. So the remote tree has picked up duplicate directories somewhere. Related to the self-conflict duplication already recorded in memory, but not yet attributed here.
Both are folder-identity problems rather than the mount disagreement originally suspected. Renamed scope accordingly; the "two mounts disagree" framing is retracted.
B14 — Provisional (ciphertext) sizes make rsync read short and abort the file
Status: Fixed (unverified — needs a cold folder on the rebuilt daemon) Found: 2026-07-20, same rsync run as B13. Alongside the receiver errors, the sender failed to read its own source files:
rsync: [sender] read errors mapping "/home/narl/ProtonDrive/Music/…/
Buunshin - i think i feel... (Original Mix).wav": No data available (61)
Where: node_size (pdfs-fuse/src/lib.rs:3480) + Core::spawn_size_upgrade
Cause: this is B12's known tradeoff surfacing as a real transfer failure.
Until the background size upgrade lands, node_size falls back to
total_size_on_storage — the ciphertext size, which is larger than the
plaintext (measured at +59 bytes on this account). B12 called that "cosmetic and
self-corrects" because it only affects ls -l.
It does not only affect ls -l. ENODATA (61) is what rsync sets on a short
read: map_ptr in fileio.c asks for the bytes stat promised, gets fewer,
and marks the mapping ENODATA. So a reader that trusts st_size — rsync,
anything using mmap, sendfile, or a sized read loop — sees a truncated
file and errors out. That is the same class of bug as B11, just from the other
direction: B11 reported too small (0) so the kernel skipped reads entirely;
this reports too large so reads run off the end.
The 30 s attr TTL makes the window much wider than the upgrade's own latency
(B12 documented this), so a cold rsync of a large tree is essentially
guaranteed to hit it.
Why it looked fine afterwards: re-listing the same directory now shows every
size correct, because the upgrade has long since landed. The bug is only visible
cold. Reproducing it needs pdfs refresh <dir> immediately before the read.
Fix: a provisional size is never published. getattr on a file whose
claimed_size is unknown now resolves it before replying, instead of replying
with the ciphertext size and upgrading afterwards.
The cost is one batched round trip per folder, not one per file: ls -l is one
getattr per entry, and they collapse onto a single upgrade. B12's split is
still doing its job — a plain ls never reaches this path at all, which is
where the 8.7× came from.
That required making the upgrade awaitable, which is most of the change:
size_upgradeswent fromHashSet<u64>toHashMap<u64, Arc<SizeUpgrade>>, a condvar per folder. Followers wait on it; there may be hundreds for one folder, so it releases all of them.- The leader does the fetch on its own thread rather than handing it to a
worker. This is the part to not undo:
getattrwaits onLane::Meta, so a leader that queued its fetch onto that same lane could have a wide enoughls -lfill the lane with threads waiting for a job that can never be scheduled. SizeUpgrade::WAITcaps the wait at 10 s, falling back to the provisional size. Astatthat never returns is worse than one that is briefly wrong.upgrade_sizesowns the single-flight bookkeeping and has one exit path;apply_size_upgradeholds the body it used to inline, so a failed fetch still releases the waiters and leaves the folder retryable.
Tests: size_upgrade_tests covers the follower being released, a waiter
arriving after finish (it checks the flag, not the notification, which it
would miss), and all waiters waking.
Not covered, worth watching: Lane::Meta can now hold waiting getattrs.
There is no deadlock — leaders never queue their own work — but a cold ls -l
across many folders at once could occupy the lane. If metadata ops start
feeling sticky under heavy cold listing, this is the first suspect.
Verified (2026-07-20) — correct, but expensive, and not completely closed
Two further holes turned up during verification; both are fixed, and the second is the one that mattered.
upgrade_sizes_for_parentearly-returned when the parent listing was not resident inchildren— and a rename invalidates exactly that, so a freshly renamed file always landed in the dead branch.statread 67 bytes for a 16-byte file, andcatreturned the 16 real bytes followed by 51 NULs. The same shape as B4: an early return that assumed the hot cache was authoritative. Now falls back to resolving the single node, keyed by its own inode.lookupreplies with attrs and a TTL too. With noreaddirplus,ls -lis onelookupper entry and zerogetattrcalls — so fixing onlygetattrfixed a pathls -lnever takes. This is worth remembering: the first fix looked right and did nothing for the reported symptom.serve_lookupnow resolves as well, taking anoff_loopflag becauselookup's warm path runs on the dispatch loop where it may not block (B5).
| measurement | before fix | after fix |
|---|---|---|
cold ls -l, 7-file wav folder |
ciphertext sizes | exact settled sizes |
renamed file stat / cat |
67 B, 51 NULs | 16 B, clean |
cold plain ls, 793 entries |
4.48 s | 4.58 s (B12 intact) |
cold ls -l, 793 entries |
~4.5 s, wrong sizes | 85.8 s, 785/793 correct |
The cost, and the part still open
A cold ls -l of a 793-entry folder went from ~4.5 s to 85.8 s. That is not
a flaw in the batching — it is B12's cause 2 arriving on schedule. A real
claimed_size requires the per-file node-key unlock, i.e. one S2K per file, and
that work is single-threaded. B12 removed it from the listing path; asking for
sizes puts it back, because sizes are what it produces. Plain ls is untouched,
which is why the split is still worth having.
8 of 793 entries still came back provisional, caught by diffing the cold
listing against a settled one. Those are SizeUpgrade::WAIT timeouts: the leader
takes ~80 s for the batch and a waiter gives up at 10 s. So the bug is rarer but
not gone — it now needs a folder large enough that the batch outruns the cap.
Per-chunk wakeup (done, unverified)
The timeout gap is closed by releasing waiters per chunk instead of once at
the end. SizeUpgrade carries a generation counter; run_size_upgrade fetches
in SIZE_UPGRADE_CHUNK (150, the SDK's own MAX_BATCH_COUNT, so one chunk is
one request), applies each chunk, and bumps the generation. A waiter re-checks
whether its own node now has a real size and returns if so — it no longer
waits on the other 792.
Two structural points worth keeping:
- The batch runs on its own thread, not the
Workerspool. Callers wait onLane::Meta, so queueing the batch there could let a widels -lfill the lane with threads waiting for a job that can never be scheduled.Lane::Transferwould swap that deadlock for starvation behind bulk reads. One short-lived thread per folder, bounded by the single-flight, avoids both. wait_forevaluates its predicate with noSizeUpgradelock held. The predicate readsstate, and the applying thread holdsstatebefore it signals; taking them in the other order would close the cycle.
This does not make a full cold ls -l faster — every size still has to be
computed. It makes each individual stat return after its own chunk rather than
after the whole folder, and removes the provisional-size fallback that the 10 s
cap was producing.
Tests: size_upgrade_tests covers the resolving chunk releasing a waiter,
an unrelated chunk not releasing one, finish releasing an unresolved waiter,
the already-resolved and post-finish arrivals, all waiters waking, and the
timeout backstop still firing.
Plan: parallelise the per-file S2K (the 85 s)
This is the only item that reduces the work rather than redistributing it, and it is the same bottleneck as improvements.md PERF #5 — approached from sizes rather than from cold navigation. Filed here because B12/B14 are what produced the measurements it needs.
The claim to be tested first, before any code. B12's profile of a 793-entry
cold enumerate attributed 64.4 % of cycles to sha2::sha256::x86::digest_blocks
and ~74 % overall to SHA-256 inside PGP's S2K, i.e. the per-file node-key
unlock in build_node under NodeDetail::Full. If that still holds for the
size upgrade path specifically — it is the same enumerate_nodes call, so it
should — then the work is CPU-bound, embarrassingly parallel across files, and
currently serialised on one thread.
Step 0 — confirm, do not assume. Re-profile against the size-upgrade path now that it is the thing on the critical path:
perf buildid-cache -a target/release/pdfs # symbols; see B12's note on strip
perf record -p <daemon pid> -- sleep 90
# meanwhile: pdfs refresh Music/aC_ID.dll && ls -l Music/aC_ID.dll
Expect the same SHA-256 leaf. If it is not dominant, stop — the rest of this plan is aimed at the wrong thing.
Step 1 — establish the ceiling. The upgrade is enumerate_nodes over 793
uids. Measure single-node unlock cost directly (log the S2K iteration count and
elapsed time per file across one folder — this also settles the question B12 left
open, whether the 49 ms/entry folder is expensive because of n or because the
uploading client chose a costlier S2K). Ideal speedup is bounded by core count
and by how uneven those per-file costs are.
Step 2 — move the crypto off the async runtime. The decryption currently runs
inline in the SDK's enumerate_nodes future. Blocking CPU work on a Tokio worker
starves everything else sharing that runtime. Wrap the per-node unlock in
spawn_blocking, or hand the whole chunk to a rayon pool and await the result.
This is the change PERF #5 describes and is a proton-drive-rs change, not a
pdfs one — which is why it was deferred: it needs an SDK release, and the
workspace pins "0.1" so a cargo update picks it up.
Step 3 — parallelise across files within a chunk. 150 nodes per chunk, each
independent once the parent key is resolved. resolve_parent_key_ctx caches the
ancestor key and all children of a folder share one parent, so the parallel
region needs read-only access to it — take the folder key once, then fan out.
Watch FOLDER_KEY_CACHE_CAP/folder_keys (PERF #9 made it an LRU) for
contention if the fan-out takes that lock per node.
Step 4 — re-measure the same three folders, which have before/after numbers already recorded above and in B12:
| folder | cold plain ls |
cold ls -l (today) |
|---|---|---|
Music/aC_ID.dll (793) |
4.58 s | 85.8 s |
InstantUpload/Camera (484) |
2.91 s | not measured |
Pictures/old (417) |
2.25 s | not measured |
Fill in the middle column's ls -l first — there is only one data point for the
regression right now, and B12's "cliff past ~500" framing was wrong precisely
because it generalised from a single folder. Do not repeat that mistake here.
What success looks like: cold ls -l on the 793-entry folder within ~2–3× of
the plain ls, rather than 19×. Perfect scaling is not available — the request
round trips are serial per chunk and only the crypto parallelises.
Explicitly out of scope: readdirplus. It would cut the number of calls
(one reply per directory instead of one lookup per entry) and is the right
long-term shape for the directory-read path, but it does not remove a single S2K
— the sizes still have to be computed. It is a separate change and should not be
bundled in, for the same reason B5 declined it: it would hide this bug rather
than fix it.
Do not attempt to derive the plaintext size from total_size_on_storage
instead. The delta is not a constant (+564 B on a 39 MB file, +51 B on a 16-byte
one — it scales with block count), so the arithmetic would have to model Proton's
block framing exactly, and being wrong in the small direction reproduces B11,
where a short size made the kernel skip reads entirely.
Verify with: the ciphertext delta is the probe (+564 B on a 39 MB file,
+51 B on a 16-byte one — it scales with block count, so it is not a constant).
pdfs refresh a folder, then ls -l immediately and diff against a settled
listing.
B13 — rename over an existing destination fails instead of replacing it
Status: Fixed (unverified — needs a real rsync against the rebuilt daemon)
Found: 2026-07-20, user ran rsync -rauLP ~/ProtonDrive/Music/... ~/Music/
(both are protondrive mounts) and every single file failed at the end of its
transfer:
rsync: [receiver] rename ".../.Buunshin - heimwee (Original Mix).wav.b0akm3"
-> ".../Buunshin - heimwee (Original Mix).wav": Input/output error (5)
Where: crates/pdfs-fuse/src/filesystem.rs, Filesystem::rename
Daemon log, one per failed file:
ERROR pdfs_fuse: rename failed uid=… error=proton api error AlreadyExists (http 422):
A file or folder with that name already exists
WARN pdfs_fuse::drain: pending upload failed; will retry uid=… attempts=1
error=proton api error DoesNotExist (http 422): File or folder not found
Cause: rename(2) is specified to atomically replace an existing
destination. Our handler just calls client.rename_node, and Proton refuses a
name that already exists — so the 422 becomes a blanket EIO. Nothing in the
path ever looks up the destination name.
This breaks every write-to-temp-then-rename tool, which is most of them: rsync,
editors doing atomic saves, mv -f, package managers. The failure is at the
end of the transfer, so the bytes are uploaded and then thrown away — the user
pays full upload cost for nothing.
Second-order damage: the failed rename leaves the temp node behind and its
queued upload then fails with DoesNotExist and retries forever. So each failed
file also leaves a poisoned entry in the drain queue.
Fix: rename now looks up newname under newparent and, if it resolves to
a different node, removes it before the API call.
The order is forced by Proton: the name has to be free before rename_node will
take it, so it is trash-then-rename, which means the operation is not atomic.
Core::remove_replaced does the trashing and Core::restore_replaced puts the
node back if the rename then fails — on every failing path, including the queued
ones. If the restore also fails the node stays in the trash and says so
loudly, because the alternative is a file the user believes was merely renamed
sitting somewhere they will not look.
Deliberate details:
- Refusals happen before anything is trashed.
check_replaceableis a pure function for exactly this reason: everyErrit returns is a case where the destination must survive, and a mistake there turns a refusal into deletion.EISDIR/ENOTDIRwhen the two ends disagree about being a directory, andENOTEMPTYfor a non-empty destination directory — Proton trashes a folder with its whole subtree, so allowing that would discard every file underneath without ever naming them. RenameFlagsis now read (it was_flags).RENAME_NOREPLACEreturnsEEXISTinstead of replacing;RENAME_EXCHANGEreturnsEINVAL, since no Proton primitive swaps two names and emulating it would leave a window in which one of them does not exist.- A node whose own create is still queued never reached the server, so replacing it just discards its queued ops — no API call, and it works offline.
- Renaming a node onto its own name stays a no-op rather than a self-replace.
Tests: replace_tests in pdfs-fuse/src/lib.rs covers all five decisions.
The trash/restore half needs a live server and is not covered.
Repro: cd <mount> && echo a > x && echo b > y && mv y x → was EIO, should
now succeed with x holding b.
B12 — Cold enumeration is slow per entry, and goes superlinear past ~500
Status: Both causes fixed and verified (2026-07-20); attr-invalidation follow-up unverified
Found: 2026-07-20, measuring improvements.md P2.8 ("pipeline
enumerate_nodes_detail") before implementing it, to decide whether the
bottleneck was network or crypto.
Measured (cold; pdfs refresh <dir> then a timed ls -1)
| entries | time | per entry |
|---|---|---|
| 1–9 | 350–550 ms | fixed cost |
| 34 | 629 ms | — |
| 147 | 1.92 s | 12.5 ms |
| 251 | 3.18 s | 12.4 ms |
| 484 | 6.25 s | 12.5 ms |
| 793 | 38.8 s | 48.9 ms |
The 793 figure reproduces to ±100 ms across three runs. Marginal cost from 484 to 793 is 106.7 ms/entry, 8.5× the baseline.
perf record against the running daemon shows ~90 % of cycles inside the
pdfs binary in both cases, and sample counts matching wall time (1022 samples
≈ 5.1 s, 7538 ≈ 37.9 s at 199 Hz) — so this is CPU-bound throughout, not waiting
on the network. Symbols are unavailable (the installed binary is stripped), which
is what blocks attribution.
Cause 1 (found, ~half the baseline): every FTS row is deleted by a full scan
nodes_fts declares uid as UNINDEXED:
CREATE VIRTUAL TABLE nodes_fts USING fts5(uid UNINDEXED, name, tokenize='trigram')
and Db::upsert_nodes (pdfs-core/src/db/nodes.rs) deletes by exactly that
column on every node, because "FTS5 has no UPSERT":
DELETE FROM nodes_fts WHERE uid = ?1
An UNINDEXED FTS5 column is not searchable, so the predicate cannot use an index.
EXPLAIN QUERY PLAN confirms it:
`--SCAN nodes_fts VIRTUAL TABLE INDEX 0:
One full scan of the FTS index per node written. Measured on a .backup copy of
the live 171 MB DB (17 443 indexed nodes):
| deletes | time | per delete |
|---|---|---|
| 484 | 3.09 s | 6.4 ms |
| 793 | 5.26 s | 6.6 ms |
So ~6.5 ms of the 12.5 ms per-entry baseline is this one statement, and it
gets worse for every user as their node count grows — the scan is over the whole
index, so the cost of writing any listing scales with the size of the account.
It is also paid by every sync pass and every event-driven refresh, not just ls.
Fix direction: delete by rowid instead. FTS5 deletes by rowid efficiently, so
map nodes.rowid to the FTS rowid and drop the uid column from the index
(or keep it UNINDEXED purely for retrieval). Needs a schema migration on a live
171 MB DB, so it wants its own change rather than being smuggled into this one.
Cause 2 (the dominant one): an S2K key derivation per file, to list a folder
Symbols came from the local unstripped build, not a new install. strip
preserves both the build id and the text addresses, so when /usr/bin/pdfs and
target/release/pdfs reported the same build id, the stripped binary was just
a copy of the local one and perf buildid-cache -a target/release/pdfs was
enough to symbolise a recording of the running daemon. Worth remembering — it
turned a "needs another install + restart" into a five-second step.
Hot functions over a 793-entry cold enumerate:
| % | symbol |
|---|---|
| 64.4 | sha2::sha256::x86::digest_blocks |
| 6.0 | <D as digest::DynDigest>::update |
| 5.7 | sqlite3VdbeExec (cause 1, above) |
| 4.0 | sha2::sha256::compress256 |
| 1.5 | pgp::types::s2k::StringToKey::derive_key |
~74 % of the listing is SHA-256 inside PGP's S2K, i.e. the per-file node-key
unlock that build_node does under NodeDetail::Full. derive_key itself shows
1.5 % because the time lands in its SHA leaf. Nothing downloads during an ls, so
PGP is the only plausible source of that SHA.
The "cliff past ~500" framing was wrong
That was an inference from a single folder, not a measurement. An S2K's cost is
set by the iteration count in the key packet, which is chosen by the client that
uploaded the file — so it varies per file, not per listing size. Every folder
measured sits at 12–16 ms/entry except Music/aC_ID.dll at 49 ms, and
Music/ELDEN RING SOUNDTRACK (67 files, same parent tree) is ~16 ms/entry, so it
is not "Music files are expensive" either. A 4× jump for a 1.6× change in n fits
"those files were uploaded by a client using a costlier S2K" far better than a
size threshold. n and folder identity are confounded in the data — there was only
ever one folder above 500 entries.
Left unresolved deliberately: the fix below removes the per-file S2K from the listing path entirely, so which of the two explanations held stopped mattering. To settle it anyway, log the S2K iteration count per file across one enumerate of each folder.
Fix: enumerate cheap, upgrade sizes in the background
ensure_children now calls enumerate_nodes_light, which skips the file
node-key unlock (folders are still unlocked — their keys are what the children
decrypt with). The listing is served from that immediately.
Light returns no claimed_size, and ls -l wants sizes, so a naive split would
just move the same S2K onto the first stat. Instead Core::spawn_size_upgrade
fetches the full nodes for that folder on a worker, batched, after the listing has
been answered, and adopts only the size — re-interning wholesale would clobber a
name or parent that a concurrent rename/move had changed. It is single-flight per
folder, because a stat of one entry in a fresh listing means a stat of all of
them. Queued writes keep their optimistic size through it, via the same
stamp_pending_sizes that B11 added.
Three entry points cover every way a provisional listing can appear: the network
enumeration, the DB fast path (rows persisted before an upgrade ran), and
getattr itself (a listing restored by hydrate on mount, which ensure_children
returns early for).
Known tradeoff — sizes are provisional until the upgrade lands. node_size
falls back to total_size_on_storage, the ciphertext size, so a stat in that
window reads slightly too large. Reads are unaffected: the revision reader
carries its own authoritative size. Deliberately not the B11 shape — that
reported 0, which made the kernel skip reads entirely; too-large is cosmetic
and self-corrects.
The window was longer than "a round trip", and that took a correction. The
daemon has the real sizes quickly, but the kernel keeps the provisional attrs
for the full 30 s entry TTL, so ls -l kept reporting them long after the DB was
right. This nearly caused a misdiagnosis during verification: two successive
ls -l runs agreed with each other and both disagreed with the truth by a
constant +59 bytes, which reads exactly like "the upgrade never ran". Querying the
DB directly is what separated "not computed" from "computed but not visible".
Closed by having spawn_size_upgrade call notifier.inval_inode for the inodes
it corrected, after the DB write so a provoked re-getattr cannot race the
persistence. That needed a Notifier on Core (a OnceLock, since the session
is built from the Core), and a fresh one per on-demand fork — each fork has
its own inode space, so notifying through the primary mount's channel would name
inodes that session has never heard of.
Verified (2026-07-20)
| folder | before | after |
|---|---|---|
Music/aC_ID.dll (793) |
38.8 s | 4.48 s (8.7×) |
InstantUpload/Camera (484) |
6.25 s | 2.91 s (2.1×) |
Pictures/old (417) |
5.10 s | 2.25 s (2.3×) |
Schema v15 migrated the live 171 MB DB on start (17 443 rows indexed). All 793
sizes converge to exactly their pre-change values, checked against a .backup
copy taken before any of this landed. The +59-byte ciphertext delta is a precise
probe for a provisional size, and is what the attr-invalidation follow-up should
be tested with.
Original cause-2 ruling-out (kept — all still true)
Backing cause 1 out of the totals leaves per-node non-FTS cost at 6.2 ms at n=484 but 42 ms at n=793. Something gets ~7× more expensive per node, sharply, somewhere around 500–800 entries.
Ruled out so far:
- Chunking / network.
MAX_BATCH_COUNT = 150, so cost would rise in visible steps at 150/300/450; it is flat at 12.5 ms/entry straight through 484. The marginal cost would imply ~16 s per POST, which is absurd. - Composition. Both the 484 and 793 folders are 100 % files, no subfolders.
FOLDER_KEY_CACHE_CAP(512). Suspicious number, butfolder_keysonly holds folder keys and these children are all files, so nothing is inserted during the walk.resolve_parent_key_ctxcaches each ancestor it derives, and all children share one parent, so it is a hit after the first child.- The entity cache.
InMemoryCacheRepositoryisHashMap-backed, O(1)set. - FTS (cause 1). Linear — 3.1 s → 5.3 s where the cliff is 6.1 s → 38.8 s.
Next step: symbol-level profile. Needs an unstripped pdfs installed and the
daemon restarted, then perf record -p <pid> across a 793-entry enumerate. Every
cheaper avenue above has been spent. Note a second daemon is not an option
for this: Proton refresh tokens are single-use, so a parallel session would fight
the running daemon for them.
Do not implement improvements.md P2.8 before this is understood. Pipelining would parallelize the cliff rather than remove it, and at 793 entries the cliff is 33 s of the 38.8 s — far more than concurrency could win back.
B11 — A file moved while its create is still queued reads as empty
Status: Fixed (verified 2026-07-20) Found: 2026-07-20, while verifying B1 on the restarted daemon.
The B1 repro is echo hi > f.txt && mv f.txt d/, i.e. the move lands while the
create is still queued for upload. Immediately after the mv, the file is
present in the destination listing — B1 is genuinely fixed — but stats as 0
bytes and cat returns nothing:
-rw-r--r-- 1 narl narl 0 Jul 20 00:29 zz-b1-dst/zz-b1-src.txt
A few seconds later the same file reads correctly (3 bytes, hi). A control
file created and read without an intervening mv was correct immediately.
Cause: State::intern_mem replaces an existing entry's node wholesale
(e.node = node). A move invalidates both parents' listings (that is B7's fix,
and it is correct), so the next ls re-enumerates — and the node that comes back,
from the remote or from its DB row, carries the size of the revision the server
holds. For a file whose write is still queued that is the pre-write size, usually
0. Interning it reverts the optimistic size record_pending_write had stamped.
Why an empty read rather than a short one: a file that stats as 0 bytes gets
no read from the kernel at all. read_range would have served the staged
blob quite happily; it is never asked. So the file reads as empty for as long as
the stale size stands, and "empty file" is indistinguishable from "file whose
contents were lost" to whatever is reading.
Core::hydrate already solved exactly this for the restart case — it stamps
each pending write's size onto the node as entries materialize. The protection
was simply missing on every live re-enumeration path.
Fix: Core::stamp_pending_sizes re-applies the optimistic size to a batch of
nodes, called from both arms of ensure_children (the DB fast path and the
network path) before the state lock is taken. It snapshots the pending map first
and returns: no site in the daemon holds pending and state at once, and this
is not the place to become the first — hydrate established that pattern for the
same reason.
Scoped to ensure_children deliberately. The other intern sites are
single-node and all authoritative by construction (mkdir, create, an upload that
just landed); drain.rs's post-upload adoption is supposed to take the
server's node, and already handles the queued case through rebaseline_pending.
Tests: pending_size_tests in pdfs-fuse/src/lib.rs covers the queued file
keeping its size, a settled sibling keeping the server's, folders being left
alone, and an empty pending map changing nothing.
Verified 2026-07-20 on the rebuilt daemon: echo hi > f.txt && mv f.txt d/
then an immediate ls -l and cat gives 3 bytes and hi, against 0 bytes and
empty output before.
B10 — GLib critical when the launcher closes over an in-flight open
Status: Open (side finding)
Found: 2026-07-19, seen in pdfs-prompt output immediately after a
successful Enter-to-open while verifying B9.
GLib-GIO-CRITICAL: g_list_store_remove: assertion '!g_sequence_iter_is_end (it)' failed
What is known: it fires right after the opening path=… log line, i.e.
during xdg_open + window.close(). It is a warning, not a crash — the open
itself succeeded.
What it is not: the launcher's own row handling. Section::set_rows walks
GtkListBox children (first_child/remove) and touches no GListStore, so
the failing store belongs to GTK/libadwaita internals, most likely something
teardown-ordering related in the app/window bookkeeping as the window is closed
while a launch is still settling.
Why it was not chased: it appeared while verifying an unrelated fix and has no user-visible effect. Worth revisiting if the launcher ever misbehaves on close (a hang, a lost open, or a stale window), since an assertion firing during teardown is exactly the shape of bug that later turns into one of those.
Where to start: run pdfs-prompt under G_DEBUG=fatal-criticals to turn
the warning into an abort and get a real backtrace naming the store.
B9 — Enter did not open the selected result in the launcher
Status: Fixed (verified)
Found: 2026-07-19, user reported Enter doing nothing in pdfs-prompt while
the arrow keys and Escape worked normally.
Where: crates/pdfs-gui/src/prompt.rs, the window key controller
Cause: key-event phase, not key handling. gtk4::Entry — really its inner
GtkText — binds Return to its own activate and consumes it. The launcher's
EventControllerKey is attached to the window in the default bubble
phase, so the focused entry saw Return first and the window handler was never
reached. Escape and Up/Down worked precisely because GtkText has no bindings
for them and they bubbled up as intended.
That asymmetry is the tell: when some keys reach a window-level controller and others silently don't, the ones that don't are being claimed by the focused widget.
Fix: Return moved off the window controller and onto
entry.connect_activate. Enter while focus is in the results list was already
handled by row_activated, so the two together cover both focus positions.
Rejected alternative: setting the window controller to
PropagationPhase::Capture. It fixes the symptom by putting the handler ahead
of the entry — but ahead of it for every keystroke, not just Return, which
puts text input and IME composition behind a handler that has no business
seeing them. The narrow fix has no such blast radius.
Verified: live. Pressing Enter logged opening path=…, the daemon hydrated
the file, and xdg_open ran.
B8 — No way to complete a CAPTCHA, so a gated sign-in is unrecoverable
Status: Implemented, partially verified — the bridge is proven, a real gate is not
Found: 2026-07-19, user hit proton api error Unknown (http 422): For security reasons, please complete CAPTCHA and had no way to answer it.
Where: SDK api.rs/http.rs/session.rs, pdfs-core/src/auth.rs,
pdfs-gui/src/app/pages/verify.rs
Cause: three gaps stacked, none of which is a bug on its own.
ResponseCodehad no9001, so the gate deserialized toUnknown— which is why the error read as an opaque failure rather than a recoverable prompt.- Nothing read
Details.HumanVerificationToken, and the client never sent thex-pm-human-verification-token{,-type}headers the retry needs. - No UI could render Proton's hosted verification page.
Fix: HumanVerification (challenge) + HumanVerificationCredential
(answer) in the SDK, header plumbing on the session-less login calls, and
ProtonApiSession::begin_verified. pdfs-core promotes a solvable gate to
Error::HumanVerificationRequired and auth::login_interactive re-runs the
login with the earned token — the retry lives in core because the gated attempt
burns its SRP handshake, which no front-end should have to know. The GUI hosts
the page in a WebKitWebView (webkit6 0.4, pairs with the existing gtk4 0.9)
and bridges the page's postMessage to a script message handler.
Deliberately not done: email/sms verification methods (the token arrives
out of band, so the webview cannot complete them — such a gate stays a plain API
error rather than opening a page the user cannot finish), HV on authenticated
endpoints (only the login path is plumbed), and re-gating of the retry (a second
challenge means the token was rejected; looping would trap the user).
Verified: the WebKit bridge end-to-end with a throwaway probe — handler
registers, the page's postMessage arrives, non-completion messages are
filtered, the completion token round-trips. Unit tests cover 9001 parsing, the
challenge/Details shape, URL escaping, and message filtering.
Not verified: a real gated login against verify.proton.me. It cannot be
triggered on demand, so the exact message shape Proton's page posts is taken
from its documented contract, not observed. If verification appears to hang with
the page solved, that is the first thing to check — extract_token in
verify.rs is the single place that decides what counts as completion.
Note: the SDK half shipped as proton-sdk / proton-drive-rs 0.1.11. The
workspace requirement was widened to "0.1" at the same time, so later 0.1.x
releases are picked up by cargo update without an edit here.
B7 — Renamed directory reads as missing until the entry TTL expires
Status: Fixed (verified 2026-07-20)
Verified: mkdir "zz-b7 Old Name" with a file inside, mv to a new name, then
an immediate ls and cat through the new name — both succeed with no wait. The
directory used to return ENOENT until the entry TTL expired.
Found: 2026-07-19, user renamed a folder on the mount with mv. ls of the
parent listed the new name, but ls <newname>/ returned ENOENT — repeatedly, so
not a one-shot race. It self-healed on its own a few minutes later, and the
files were all intact: this is a visibility bug, not data loss.
Where: crates/pdfs-fuse/src/state.rs, State::relocate
Cause: the online rename path ended with relocate, which called forget on
the moved node. forget drops the node's by_uid mapping, so when the
invalidated parent listing re-enumerated, intern_mem allocated it a fresh
inode. But the kernel had already carried the renamed dentry over to the old
inode number, and it holds that dentry for the entry TTL. Every lookup, getattr
and opendir through it resolved to an inode entries no longer held:
entries.get(old_ino) -> None -> ENOENT
readdir of the parent went the other way — it walks children and reports the
new inode — which is exactly why the directory listed fine but could not be
entered. Once the TTL expired the kernel re-looked-up the name, got the new
inode, and everything worked again.
Fix: relocate now rewrites the node in place (rename_in_place) and
invalidates both parents' listings, instead of forgetting it. The inode is
preserved, so the kernel's dentry stays valid and the re-enumeration reuses the
same by_uid slot. Both relocate tests now assert inode stability; that
property was untested, which is how this got through the B1 fix.
Note: dropping the forget also stops the moved node's DB row from being
deleted and re-created on every move — the row is updated instead, which is what
the B1 fix was working around from the other side.
B6 — Daemon sets no file modes: control.sock is an unguarded authority
Status: Fixed (verified 2026-07-20)
Found: 2026-07-19, while writing the plaintext-at-rest threat model in
docs/ARCHITECTURE.md §8. Not from a report — from checking a claim before
asserting it in a doc. I had written "their default 0700 permissions protect
them", went to verify, and found we set no modes at all.
Where: crates/pdfs-fuse/src/mount.rs (UnixListener::bind(control_socket)),
crates/pdfs-core/src/config.rs (create_dir_all on state/cache dirs)
grep -rn "set_permissions\|from_mode" crates/pdfs-core/src crates/pdfs-fuse/src
returns exactly one hit, and it is shell.rs setting 0755 on a generated
script. Nothing else sets a mode. So:
- state and cache directories are created at
0777 & ~umask— typically0755 control.sockis bound at the same, typically0755
Anything that can connect to the socket drives the daemon with its authenticated session: enumerate the tree, read file contents, upload, trash, create public share links. No credential is required — the keyring is never consulted, because the daemon already holds the session.
The only thing preventing this today is that ~/.cache and ~/.local/state are
conventionally 0700. That is a property of the user's system, not something we
establish or check. It does not hold on a machine with a permissive umask, a
group-shared home, or a home restored from a backup that flattened modes.
Severity: low on a single-user desktop, real on a shared or multi-user host. It is a privilege boundary rather than a data-at-rest issue, which is what makes it worth more than the cache-plaintext point it was found next to.
Fix: chmod 0600 on the socket immediately after bind (before the listener
thread starts accepting), and 0700 on the state and cache directories at
creation. Both are a few lines. Worth also asserting the socket mode in a test,
since a regression here is silent.
Fixed 2026-07-19: AppDirs::ensure now sets 0700 on the state, cache, and
config directories on every start (not just at creation, so an existing
permissive directory is tightened), and config::restrict_socket sets 0600 on
both the control socket and the tray socket immediately after bind. A socket
whose mode cannot be set takes the daemon down rather than serving unguarded;
the tray's is best-effort, since it only guards single-instancing. Unit tests in
config.rs assert both modes.
Verified 2026-07-20 on the restarted daemon — all three directories now read
drwx------ and both sockets srw-------, against drwxr-xr-x / srwxr-xr-x
before.
Measured exposure at the time of the fix: the directories really were
drwxr-xr-x and the live socket srwxr-xr-x, but ~/.cache and ~/.local
were both 0700 on this machine, so nothing was reachable in practice. The bug
was a latent dependency on those parents, not a live hole. Removing the
dependency is the point.
B5 — ls -l costs a network round trip per file (thumbnail xattr probes)
Status: Fixed (verified 2026-07-20)
Found: 2026-07-19, investigating "exa -l is slow in the mounts, exa is fast"
Reproduced (2026-07-19)
All runs cold (35 s wait to clear the 30 s attr TTL). exa needs only readdir;
exa -l stats every entry, so the delta is per-entry metadata cost:
| mount | entries | exa |
exa -l |
delta per entry |
|---|---|---|---|---|
primary ~/ProtonDrive/Installer |
34 | 10 ms | 11 ms | 0.03 ms |
on-demand ~/Documents |
45 | 27 ms | 54 ms | 0.6 ms |
on-demand ~/Videos/[Reaktor] FMA … |
65 | 73 ms | 193 ms | 1.85 ms |
Cause (2026-07-20): a network round trip per file per xattr name
strace -f -T on a cold exa -l of the 65-file .mkv directory:
lgetxattr(".../E01 ....mkv", "user.proton.thumbnail", NULL, 0) = -1 ENODATA <0.186435>
129 lgetxattr and 258 llistxattr calls for 65 entries, each lgetxattr
~186 ms. Three things compounded:
listxattradvertiseduser.proton.thumbnail+user.proton.previewfor every file, regardless of whether that file could have one. An xattr-aware lister then asks for each advertised name — twogetxattrper entry.Core::thumbnailcached only success.download_thumbnailreturningNone— the normal answer for a.mkv— was never remembered, so every listing re-asked the API and was re-told nothing, forever.getxattrran inline on fuser's dispatch loop, not on theWorkerspool.lookupandreaddirhad been moved off it (PERF #1.0); this one was missed. So each 186 ms miss stalled every other op on the mount, which is why the concurrencyexadoes have bought nothing.
The mount-kind correlation was a red herring. It tracks file type, not the
fork: ~/Videos is .mkv (Proton generates no thumbnail — always a miss),
~/ProtonDrive/Installer is installers whose probes were already warm. The
fork_state theory in the previous write-up is wrong; forked and primary mounts
run the same code and neither is at fault. Worth noting as a lesson — three
measurements on different directories looked like a mount-kind effect because
nobody had varied file type independently.
Fix
listxattradvertises the names only forimage/*andvideo/*media types.getxattrstill honours an explicit request for an unadvertised name, so nothing becomes unreachable. Everything else — documents, installers, archives — now costs zero round trips per listing.Core::thumbnailremembers misses inno_thumbnail, keyed(uid, type)and validated against the node's mtime exactly as the positive side is. Bounded by clearing atMAX_THUMBNAIL_MISSES(8192).getxattrhands off toLane::Meta, so a miss can never block the dispatch loop.
A video directory still pays its misses once per revision (image/video is where a thumbnail plausibly exists, so the probe is legitimate) — but in parallel, and never again.
Verified (2026-07-20)
Cold runs (40 s wait each) on the restarted daemon:
| measurement | before | after |
|---|---|---|
~/Videos/…FMA… (65 .mkv) cold exa -l |
193 ms | 14 ms |
~/Documents (45 entries) cold exa -l |
54 ms | 13 ms |
user.proton.* probes over the 65 .mkv |
130 | 0 |
~/Documents probe latency, 1st listing |
124–395 ms | 124–395 ms |
~/Documents probe latency, 2nd listing |
124–395 ms | 0.26–0.67 ms |
Each part of the fix is separately visible in the traces:
- The advertising gate. The
.mkvdirectory now issues zerouser.proton.*probes; the only remaininglgetxattris exa's ownsecurity.selinuxat ~0.3 ms. Note this means Proton does not report those files asvideo/*— the gate is excluding them by media type, not because they are known to lack a thumbnail. A video whose media type is generic no longer advertises one even if it has it;getxattrstill serves an explicit request, so it stays reachable. Worth revisiting if thumbnails ever go missing in a file manager. - The negative cache.
~/Documentsholds 10.pngfiles, correctly still advertised, so it probes 20 times on both passes — but the second pass answers in 0.26–0.67 ms instead of going to the wire. That is the whole point: the probes that are legitimate stop being expensive after the first. - The worker handoff. Pass 1 on
~/Documentsshows interleaved<... lgetxattr resumed>lines across several tids, i.e. the misses now overlap instead of serializing behind the dispatch loop.
Secondary finding (still open): lookup is O(n) per name
Where: crates/pdfs-fuse/src/filesystem.rs, serve_lookup
serve_lookup linear-scans the parent's children comparing names, and with no
readdirplus that is one lookup per child — O(n²) name comparisons under the
global state lock for a listing.
Real, but not what made this slow: it costs the same on both mount kinds and at 65 entries a linear scan is nanoseconds against a 186 ms round trip. Two candidate optimizations were considered and deliberately not applied, and that judgement still holds now the real cause is known:
- Per-directory
name -> inomap. A second structure that must stay in sync withchildren— the two-halves-of-one-cache shape that caused B1 and B4. readdirplus(fuser0.17 supports it; we returnENOSYS). The right shape long-term, but it is a rewrite of the directory-read path and it would have hidden this bug rather than explained it.
Measurement notes for next time
- Cold requires a 35 s wait per run (30 s attr TTL). Warm runs show ~6 ms regardless and prove nothing.
strace -conexaneeds-f— exa is multithreaded, and without it the trace captures only main-thread startup (3statxcalls, nogetdents).-T(per-call durations) is what cracked this;-csummaries attribute time across threads in a way that hid the 186 ms constant.
B4 — invalidate_listing silently skipped non-resident folders
Status: Fixed (verified 2026-07-20, via B1's repro — the FUSE rename path that
depends on this helper now behaves correctly end to end)
Found: 2026-07-19, reviewing the B1 fix
Where: crates/pdfs-fuse/src/state.rs:257
if self.children.remove(&ino).is_none() {
return; // never clears the DB `listed` flag
}
The early return assumed "not in the hot cache ⇒ nothing to invalidate", which
stopped being true once listings became DB-backed. A folder trimmed from the
in-memory map but still listed = 1 in the DB could not be invalidated at all —
callers thought they had dropped the listing, and ensure_children would happily
rebuild the stale one.
Fix: drop the early return; always clear the flag. Costs one redundant
UPDATE when nothing was cached.
Core::refresh_dir had been hand-rolling a workaround for exactly this (clearing
the DB flag itself, then reaching into state.children directly, with a comment
explaining why it couldn't use the helper). It now just calls
invalidate_listing. Behaviour change worth knowing: refresh_dir used to
propagate a DB write failure to the caller of CtlRequest::Refresh; it now warns
and reports success, matching every other invalidation site.
Test: a_deleted_child_leaves_a_listed_parent_serving_a_stale_listing in
pdfs-core/src/db/tests.rs pins the B1/B4 mechanism at the DB layer — a deleted
child plus a still-listed parent yields a listing that silently omits it.
B3 — Activity log timestamps written in seconds, read as milliseconds
Status: Not a bug — investigator error, kept as a record Found / retracted: 2026-07-19
Originally filed because every activity row rendered as 1970-01-21 …. That was
an artifact of the ad-hoc debugging query, which divided by 1000; the code never
does. Seconds are consistent end to end:
- writer —
log_activityusesnow_secs()(pdfs-fuse/src/lib.rs) - schema /
activity_list— pass the value through untouched - reader —
activity_timecallsglib::DateTime::from_unix_local(secs)
The correct manual query, for next time:
select datetime(time,'unixepoch','localtime'), kind, target, detail, ok
from activity order by time desc limit 40;
Worth noting because the bad query is what made the activity log look empty
around the reported mv — which is evidence B2 leans on. That absence has since
been re-checked with correct scaling and it does hold: no row at 15:58.
B2 — The reported .mkv went to trash, the repro vanished entirely
Status: Partly resolved — mv exonerated; trash origin still unattributed
Found: 2026-07-19, while chasing B1
What was settled (2026-07-19)
mv never trashes. Three repros on the mount, all with the old (pre-B1-fix)
daemon:
| repro | node state | destination | result |
|---|---|---|---|
zz-claude-repro-src.txt |
create still queued | fresh folder | vanished, not trashed |
zz-b2-settled.txt |
settled, real uid, landed | listed folder | vanished, not trashed |
zz-b3 Movie Title.mkv |
settled | folder named as the file's stem (the user's exact shape) | vanished, not trashed |
So B1 fully explains the disappearance, and nothing in the rename path
trashes. The .mkv reaching the trash has a different cause.
The listing the user acted on was almost certainly stale. B1/B4 make stale
listings persistent — listed = 1 lives in the DB and survives restarts — so
ls showing the file at 15:57 is not evidence it still existed remotely. This
is the most likely reading: it had been trashed earlier and the mount kept
showing it.
Ruled out:
- Sync engine.
reconcile_foldergates onmode == "mirror";~/Videoswasondemandfrom 15:39:27. All three sync trash sites log, and no such rows exist. - Log pruning hiding the evidence.
ACTIVITY_KEEP = 2000; the table is at the cap but its oldest row is Jul 17 18:03, so the whole window is covered. The.mkv's uid appears exactly once — anUploadat 06:07:56 — and never again. - Conflict machinery.
keep_as_conflict_copyuploads a new file under an alternate name; it never trashes the original. - Eviction through the mount. Was the best theory:
mirror→ondemandcallsevict_dir_contents, which would issue oneunlink(2)per file, and FUSEunlinktrashes remotely. Butapply_sync_folder_modeevicts beforespawn_ondemand_mount, so the deletes hit the plain local directory. Dead. - Control-socket delete.
CtlRequest::Deletelogs on both success and failure.
What remains: trash_child (backing FUSE unlink/rmdir) was the only
trash path that wrote no activity row — so an ordinary rm on the mount, at any
point after 15:39:27 when ~/Videos became a FUSE mount, would produce exactly
what we see and leave no trace. Note there is also a trashed folder of the same
name (Evangelion 1.11 - You Are (Not) Alone, is_dir=1, content mtime
04:57:50), which suggests this mkdir-and-move dance had been attempted earlier —
plausibly followed by a manual cleanup while B1 was making files appear to vanish.
That is a hypothesis, not a finding. It is not provable from the evidence that survives.
Fix applied
trash_child now logs to the activity feed on both success ("trashed from the mount") and failure, matching every other trash site. The next occurrence will
be attributable — which is the part that actually mattered here.
If it recurs
The activity log is now the first place to look. trash rows carry no
trashed-at timestamp (the mtime column is the node's content mtime — for the
.mkv it read 06:07:56, matching its upload, not its deletion), so the activity
feed is the only ordering evidence there is.
Two different endings for what looked like the same operation, so B1 may not be the whole story:
- User's
Evangelion 1.11 - You Are (Not) Alone.mkv→ trash, full size (1768670449) intact, recoverable viapdfs restore. - Clean repro (
zz-claude-repro-src.txt) → gone entirely. Not in trash, nonodesrow, nothing in the search index.
B1 explains the second. It does not explain a node reaching the trash — nothing
in the FUSE rename path calls trash_nodes.
Leading theory: the trash came from the sync-conflict machinery, which is
logged (ActivityKind::Trash rows exist with (sync-conflict <ts>) details).
The ~/Videos folder is registered as a sync folder in ondemand mode, and had
just been switched to ondemand shortly before. reconcile_folder gates on
mode == "mirror", so it should have been inert — worth confirming that gate
actually held, and that no pass was already in flight across the switch.
Note the file was a conflict-copy sibling of two files already carrying
-003 / -004 suffixes, so the conflict path had definitely been active in that
directory.
Next step: find the trash event's origin. The activity log had no row at the 15:58 mv, which points away from a logged (control-socket / sync) path — but the timestamp bug in B3 made that table hard to read, so re-check with correct scaling before trusting the absence.
B1 — FUSE rename loses the file (data loss)
Status: Fixed (verified 2026-07-20)
Found: 2026-07-19, user reported mv file.mkv dir/ on the mount deleted the file
Verified: the exact repro below on the new daemon — mv returns 0, the file is
present in the destination listing and reads back its content. Previously it was in
neither directory.
Where: crates/pdfs-fuse/src/filesystem.rs, Filesystem::rename
rename ended with:
st.forget(&uid);
st.children.remove(&newparent); // memory only
forget() deletes the node's DB row (db.delete_node). children.remove()
only drops the in-memory listing, leaving nodes.listed = 1 on the
destination folder. The next ls goes through ensure_children, which takes the
DB fast path (children_if_listed, lib.rs:722) and rebuilds the destination from
the database — where the node no longer has a row.
Net effect: file gone from the source, absent from the destination, and
rename(2) returned 0. Silent loss with a success code.
Severity is worse than "stale cache": listed is a DB column, so the bad
state survives a daemon restart. Nothing clears it on its own — the affected
folder keeps serving the listing that omits the file indefinitely, until an event
invalidation or a manual refresh happens to hit it. Confirmed on the repro
folders, which still read listed=1 with no child rows minutes later.
The data itself is believed intact server-side (the move succeeded; only the local view is wrong), but that was not verifiable without a daemon restart — see the status line.
Also hits plain in-place renames (mv a.txt b.txt), same mechanism — the parent
is both source and destination there.
The control-socket Core::move_to path was always correct; it used
invalidate_listing, which clears the DB flag as well. The FUSE path didn't.
Fix: use the same helper.
st.forget(&uid);
st.invalidate_listing(newparent);
Repro:
cd <mount> && mkdir d && echo hi > f.txt && mv f.txt d/
# renameat2(...) = 0, but f.txt is in neither directory,
# not in trash, and absent from the `nodes` table
Note on invalidate_listing: it early-returns when the folder isn't in the
in-memory children map, so it won't clear a stale DB listed flag in that
case. Safe on this path (both lookup_child and the explicit
ensure_children(newparent) guarantee the listing is resident), but it's a sharp
edge for other callers. See B4.
Acceptance harness findings
Defects in scripts/fuse-acceptance.py itself, not in the client. Numbered separately.
T1 — Managed acceptance raced daemon startup
Status: Fixed (2026-07-22)
Found: repeated install/restart/acceptance loop
Where: scripts/fuse-acceptance.py, managed setup
Repro: Run systemctl --user restart proton-drive.service and immediately
start --managed-live. The offline reference completed before the daemon had
recreated its control socket, so the first pdfs --json sync list failed with
ENOENT and setup aborted.
Fix: Managed validation polls the supported sync list API until the daemon
is ready or PDFS_ACCEPTANCE_SYNC_TIMEOUT expires. This is tracked as a test
harness defect rather than an application bug.
T2 — Acceptance cleanup raced queued namespace replay
Status: Fixed (2026-07-22)
Found: managed on-demand/on-demand matrix after every functional test passed
Where: scripts/fuse-acceptance.py, per-contract cleanup
Repro: A combined move+rename was correctly accepted into the durable queue.
Cleanup started recursively deleting its test tree while the rename drain was
still landing. The moved entry appeared in dir-b after rmtree enumerated it
but before rmdir(dir-b), producing ENOTEMPTY.
Fix: Before cleanup, poll pdfs --json status until both pending counters
are zero. Retry rmtree only for ENOTEMPTY inside the uniquely owned test
root, then wait for the queued deletions before changing modes. Other errors are
still immediate failures.
T3 — Mode-matrix preservation check raced mirror restoration
Status: Fixed (2026-07-22)
Found: managed on-demand/on-demand → on-demand/mirror transition
Where: scripts/fuse-acceptance.py, managed mode switching
Repro: Switch a populated managed folder from ondemand to mirror, wait
until sync list reports mirror/idle, and immediately read a preservation file.
The file can still be absent even though the subsequent restore pass downloads
it correctly.
Cause: The application commits the new mode before scheduling reconciliation and leaves the previous idle state and sync timestamp in the row. Those values describe the completed on-demand state, not a completed mirror restoration.
Fix: The harness records every transition to mirror and explicitly requests and waits for a newer completed sync pass before validating local bytes. This is tracked as a harness synchronization defect; transient absence before that pass is expected asynchronous behavior, not data loss.
T4 — Successful managed cleanup left its own sentinel behind
Status: Fixed (2026-07-22)
Found: second consecutive managed acceptance run
Where: scripts/fuse-acceptance.py, managed cleanup
Repro: Complete the mode matrix in mirror/mirror, then immediately start a new managed run. Registration cleanup removed the sync-folder records and remote test folders, but the next run's empty-directory precheck found the harness's 128 KiB preservation sentinel in each local directory.
Cause: Removing a mirror registration intentionally preserves the local copy. The harness treated unregistering as if it also emptied local storage.
Fix: After successful unregister, cleanup deletes only the sentinel whose contents match the per-run bytes generated by this harness. Unknown files remain untouched and continue to fail the strict precheck.