A listing that failed anywhere in the remote walk ended the whole pass,
so a first sync of a large folder on a flaky link walked the tree again
and again and never downloaded anything (B271, issue #46). A failed
listing below the root now records that folder as a gap: nothing under
it is deleted, the rest of the pass applies, and the report names it.
A restore that skipped every folder was logged as a successful activity
(B269). The daemon now gets the restored and skipped folders back, logs
any skip as a failed activity with the reasons, and sends the Rewatch
for folders already restored when the database insert fails partway.
`pdfs devices join` points to `pdfs activity` for the outcome.
diagnose dropped most of the Status reply and never looked at the synced
folders, so it said "No problems found" while a folder would not sync
(B268, issue #53). It now warns about writes that keep failing, a write
a day old, a pause, folders in error or conflict, and a control socket
nothing listens on.
from_api took every error that was not a ProtonError for a transport
failure. Pin, unpin, the searches and node loads hand it pdfs_core
errors from the cache and the database, so a failure there said
"You're offline" (B266, the rest of B205). The source chain now
decides, and only an error with neither type reads as offline.
A mount that failed on every try retried every 5 s forever, and the
reason was only in the journal: the app and the tray said Connecting...
for days (B263, issue #40). The retry now doubles to a 2-minute ceiling,
and the last failure is kept in mount-failure.json for the app, the
tray, pdfs diagnose and systemctl status.
The early returns in mount left the workers they had started running
beside the next attempt (B264). They are now stopped on every error
path, without telling systemd the daemon is stopping.
A new sync folder row starts with state idle, so the app called a folder
up to date before any of it reached Drive (B261, issue #53). A mirror
folder that has never synced now reads "waiting for first sync", and
the This computer row counts such folders. All catalogs translated.
A poll in the same burst as a new folder's Reconcile(id) took its place,
so the new folder waited behind a pass over every other one (B260).
A recursive watch that failed partway left its watches registered, was
never retried, and its events ran passes over every folder (B262). It
is now unwatched and retried every 10 minutes, symlinks are not followed,
and watcher errors are logged. A folder removed and added again in one
burst keeps its watch under the new id (B272).
One thread runs every synced folder's passes. A panic in a pass ended
it, and every later add, sync-now, file event and poll went nowhere
without a word until a restart, which fits issue #53 (B259).
Each pass now runs under catch_unwind; a panic sets the folder to error
and the next poll tries again. Diagnostics reports the engine's state,
the supervisor warns when it has ended or a pass has gone quiet, and
first passes log at info.
journald stored every daemon line at info, so journalctl -p warning,
which TROUBLESHOOTING.md asks reporters for, was always empty (B258).
When fd 1 is the journal stream, prefix lines with <N> and drop colours.
B258 to B273: why a stuck sync, a failing mount and a silent restore
are invisible, plus the sync, mount and restore bugs behind them.
B165, B202, B205 and B210 point at the entries that finish them.
A switch asked for before any pass saw the folder empty found it still
idle, so the check for what only this computer has ran at once and
refused it for good: every synced path read as "deleted here but not on
Proton Drive yet". The account run hit it in the B255 case.
A local copy gone in full is the wipe guard's case, which the push pass
of a queued switch settles without trashing anything on Drive. The
switch now returns NotNow for it, so the request is queued and that pass
runs first. A partial deletion is still refused as before.
A mirror folder whose files were all gone from this computer stayed in
error for good. The wipe guard rightly refuses to trash the whole folder
on Drive, but nothing got the folder out of that state: a queued switch
to on-demand waits for a pass that ends idle, and its push pass hit the
same guard. Only removing the folder from sync cleared the baseline that
made the empty folder read as "delete everything".
- The push pass for a switch to on-demand forgets the baseline rows of a
local copy that is gone in full instead of trashing them, so the
switch goes ahead and the mount shows what Drive has.
- New Request::ResetSyncFolder, `pdfs sync reset ID` and "Sync From
Scratch…" in a synced folder's menu clear a mirror folder's baseline
(and a queued switch to on-demand, whose push pass would upload a
duplicate of every file). The next pass compares both sides as on the
first sync: one-sided files are copied across, files that differ keep
both versions, nothing is deleted.
- The wipe guards' messages name the reset; restore logs why it skipped
a folder.
Two simulations cover it. The on-demand one failed in exactly the
reported state (mirror, pending ondemand, error) before the fix.
All 8 catalogs translated.
close(2) returns before the kernel sends release, and release takes the
write handle out of the open writes a few milliseconds before the
revision is in the queue. A `pdfs copy` checked in that gap found the
edit nowhere and copied the version Drive had. A folder copy looked only
at the queue, so it also missed files below it still open with unsaved
writes.
release now keeps the inode in State::releasing until the revision is
queued. A copy refuses a file or folder with unsaved writes, or a
release still queueing, at it or below it, and a move checks the same
set.
A copy inside the mount went through the kernel's read/write loop, one
FUSE round trip per page, because copy_file_range answered ENOSYS. The
daemon now copies a range whose source bytes are local (an open write
handle, a queued write, the cache) into the destination's scratch file,
the same way write stages them. A range it would have to download gets
EOPNOTSUPP, which makes the kernel copy that one call itself and keeps
the operation enabled. The copy still uploads: Drive cannot copy into a
file cp has already created.
In the simulation, a warm cp of a 256 MiB file went from about
100 MiB/s to about 450 MiB/s (debug build).
A mirror upload that Drive did not list yet when it was read back left a
baseline row with no revision id. The next walk reused the recorded size
whenever the modification time matched, and a row without an id never
counts as moved. So a revision another computer saved in the same second
looked unchanged and never arrived (B254). The B250 simulation hit this
about one run in five on CI.
A row without an id now reuses its size only while Drive gives no id
either. Otherwise the walk reads the size, and an unchanged file gets
Drive's revision id written into its row, so each row costs one extra
read. Rows from before schema 44 are read once after upgrading.
The app read the session with `auth::load().ok()`, so a locked keyring, a
Secret Service that was not up yet or any D-Bus error showed the sign-in
form as if nobody were signed in. The session was still saved (B253).
Keyring errors are now kept apart from "not signed in". The login page
shows a banner saying the saved sign-in could not be read, with a Retry
button. The app tries again after 2, 5, 10 and 30 seconds and opens on
Files once the keyring answers. The login page's header also loses the
back button the collapsed split gave it.
Translated into every shipped language.
Add B248 to B252 to the bug ledger, the shared-device feature and schema 44
to the 3.5.0 changelog, and describe the join flow, per-machine profiles,
conflict copy names and the mirror's multi-writer rules in the user guide,
recovery guide, CLI and configuration references and the architecture notes.
A Takeout photo with no sidecar and no date in its name was filed under
the day of the import, and the app dated uploads by modification time.
A new `exiftime` module reads DateTimeOriginal (then DateTimeDigitized)
with its offset. Takeout uses it for undated photos it copies out for a
thumbnail, and the daemon uses it for app uploads before falling back to
the modification time (B244).
Opening a file or a location stat'ed the mount on the GTK main thread, so
a hung mount hung the app. The browser now uses the launcher's bounded
worker check (`present_within`, 2 s), and opening a location runs on a
thread. Every child the GUI starts is also waited for on a thread, so it
no longer stays a zombie (B243).
The viewer decoded each full-size photo, read its EXIF and stat'ed it on
the GTK main thread, freezing the app for a large photo. That now runs in
gio::spawn_blocking, and a result for a photo the user has left is
dropped (B242).
The viewer's handlers held strong references to the viewer and its
window, which own them, so every closed viewer stayed in memory with its
full-size photo. The share dialog's row handlers had the same cycle with
the dialog state. The handlers now hold weak references, and the viewer's
one strong handle is dropped on close with the photo (B241).
Only `pdfs daemon` marked itself non-dumpable, though `mount`, `login`,
`unlock` and `logout` hold the same passphrases or tokens. All five now
do. The file-manager script also passes each path after `--`, so a file
named like an option stays a path (B240).
`import-google-photos --wait` and `redate-photos --wait` exited 0 however
the run ended, and a run that failed as a whole left no summary, so the
CLI said no import had run and the app said it finished. Failed runs now
leave a summary with the reason, `--wait` exits 1 on failure, cancel or
failed photos, and the app shows the reason (B239).
The name checks the drain makes before each create, rename and trash,
and the wake of ops waiting for a name, read the whole queue, which made
a large offline batch quadratic to drain. Schema 43 adds an index on
pending_op(kind, name) (B238).
A reconnect made every failed op due at once, quota and access refusals
included, and ops that failed together retried in the same second. The
doubling also stopped at 128 s, short of the 300 s maximum. Reconnects
now move up only unrefused ops, and the backoff reaches its maximum with
a per-op spread of up to a fifth (B237).
Takeout imports, re-dates and the app's photo upload all sent an empty
thumbnail list, so the photos showed blank in Proton's apps. Re-dates and
app uploads now make thumbnails from their file. A Takeout photo that can
have one is copied out of the archive to the private staging folder,
thumbnailed and uploaded from the copy (B236).
Each photo was ranked against every sidecar in its folder, so a folder
of 10,000 photos took about 30 s to scan. The sidecars are now indexed
by stem and each photo is ranked only against the stems a few range
lookups return: 0.1 s for the same folder, with the same matches (B235).
Nothing stopped `/`, the home folder, the Proton Drive mount, the app's
own state or cache, or a folder nested in another root from becoming a
sync root, and a restore took a path from another machine's profile as
it was. One validator, `syncroot::problem`, now runs where roots are
added and restored, on paths resolved through symlinks with `..`
refused (B234).
Logout only deleted the keyring entry, so the session's tokens stayed
valid on Proton's side. It now ends the session first (DELETE auth/v4,
bounded to 10 s) and signs out locally either way. `pdfs logout` says
when Proton could not be reached (B233).
Rows, toasts and status page descriptions parse their text as markup by
default. Shared item names, inviter addresses, device names, link URLs,
paths, EXIF values and daemon errors went in unescaped, so a name with
`&` showed as an empty row and other names were drawn as markup. Those
rows and toasts now use plain text, and status page descriptions are
escaped (B232).
The upload deadline assumed at least 32 KiB/s. Under an upload cap shared
by every drain worker and sync task, a large file ran out of time on every
attempt, took the mount offline and restarted from byte 0. The deadline
now counts the upload's share of the cap as the slowest rate when that
share is lower (B231).
- A gap-filled revision blob is recorded as complete once filled, so a
retry after a lost upload answer recognises its own bytes on Drive
instead of forking a conflict copy (B228).
- Commits that accept a change into the queue run at synchronous=FULL,
so an op `close()` reported as done survives a power cut (B229).
- A queued create's adoption window starts at its first send, not when
it was queued, so a create queued offline no longer takes another
device's empty node of the same name. A send record that does not
parse is kept instead of overwritten (B230).
A read on a write handle was answered on the dispatch loop to keep its
order with the handle's writes, and when the bytes it asked for were
neither written nor cached it downloaded them there, holding every request
of the mount for up to the read deadline. The loop now answers from the
scratch file and the local caches only, and hands a read that needs a
download to a transfer worker with the handle as it was when the read came
in (B227).
Thumbnails were decoded without the EXIF orientation, so a portrait photo
stored as landscape pixels plus a tag got a sideways tile, and the drain
uploaded that thumbnail to Drive for every client to show (B226).
The prefix rule for sidecars whose names Google cut at 51 characters also
matched names that were never cut, so IMG_1.jpg's sidecar gave its title
and date to an IMG_12.jpg that had none. It now applies only to a sidecar
name at the cap. Edited copies, which have no sidecar of their own, take
their original's date by a rule of their own and keep their own name
(B225).
load_config gives defaults for a file it cannot read, and every setter
loads, changes one field and saves. The first settings change after the file
was damaged therefore wrote defaults over the mountpoint, the adopted device
and the ignore rules. save_config now refuses while the file on disk does
not parse, and the first-run default is written atomically with mode 0600
(B223).
flush answered ok at once and release synced the scratch file, but close(2)
only returns what flush answers, so a full disk at close was reported to
nobody. flush now syncs a handle with writes on a worker and returns ENOSPC,
EDQUOT or EIO; release then finds the pages on disk instead of writing a
whole file back on the dispatch loop. fsync on a descriptor opened for
reading no longer fails with EBADF (B222).
fallocate passed any mode through and tracked only a punched hole or
growth: a zeroed range came back from the remote base at commit, and a
collapse or insert corrupted the file. Unsupported modes now get EOPNOTSUPP,
zeroed ranges count as written, the real errno is returned, and the state
lock is not held across the syscall (B224).
A cold listing allocated the local id of every new child in a write
transaction while it held the global state lock, then marked the folder
listed, still under it. Every inline handler of every mount waited on that
transaction, and the transaction on the write connection the drain shares.
The ids are now allocated before the lock is taken, and the folder is marked
listed through the outbox after its children's rows, so a crash in between
no longer leaves it recorded as complete and empty (B221).
A shared folder's listing is fetched again once it is a minute old, and that
fetch ran on the dispatch loop whenever the old listing was still resident,
so every request on the mount waited for Drive (B219).
Thumbnail downloads, size upgrades and the Shared with me listing had no
deadline of their own. On a stalled link each held a mount thread for the
SDK's retry budget, and a folder of photos filled the three metadata workers
that every cold lookup and rename needs. Offline, an uncached thumbnail is
now answered as absent, and these calls give up after 20 or 60 seconds
(B220).
PRAGMA integrity_check and the row counts behind pdfs cache inspect read
the whole file, 34 s on a 567 MiB database, and they held the write
connection while they did. Every write in the daemon waited behind them.
They now read a WAL snapshot on a pooled read connection (B218).
FTS5 keeps deleted entries until their segments are merged, which with
these indexes is close to never. On one machine local_fts held 317 MiB for
40,000 paths, against 20 MiB rebuilt, and search ran four times slower. The
indexes are now compacted once a day after a local scan. A scan that hit the
entry cap or lost a batch no longer prunes the paths it never reached, which
the next scan used to add back (B217).
nvrl