* [Refactor] Move the remaining zone commands out of game_graphics/zones/
* [Game] Move zone dump request onto ZoneViewZoneLogic per review
---------
Co-authored-by: Lukas Brübach <Bruebach.Lukas@bdosecurity.de>
* [Refactor] Move table row mapping from TableZone into TableZoneLogic
* [Game] Use TableZoneLogic::ROW_COUNT directly per review
* [Format] Match clang-format line breaking in table_zone
---------
Co-authored-by: Lukas Brübach <Bruebach.Lukas@bdosecurity.de>
Folds in the closed#7410 (Move AnnotationDialog from game/ logic to
game_graphics/) now that layers 1-2 are closed and this stack bottoms on
master directly:
- Add game_graphics/dialogs/dlg_annotation.{h,cpp} and move the
QInputDialog subclass there; game/ no longer declares any widget type.
- Add game/player/token_info.h and game/player/move_top_cards_until_options.h.
- Move CardMenuActionType to game/player/card_menu_action_type.h.
- Graphics dialog headers include the new game/ headers instead of
defining the structs themselves.
No behavior change.
Co-authored-by: Lukas Brübach <Bruebach.Lukas@bdosecurity.de>
* [Installer] Stop calling the payload free while an executable is locked
The guard answered "nothing is holding the payload" from two lookups that can
both miss a running instance: Restart Manager does not reliably report a
program's own image or the modules it mapped, and the image path of a process
running under another account cannot be read from the probe at all. A user log
showed exactly that: the first guard passed while cockatrice.exe was running
from the directory, the old uninstaller silently skipped the file it could not
delete, and the update aborted half an uninstall later with "cancelled by user"
after a close attempt the running window ignored.
Ask the file itself. A running image is opened without write sharing, so the
write open replacing it needs is refused with ERROR_SHARING_VIOLATION for as
long as the program runs, while a file that is merely being read still accepts
it. When neither lookup named anybody and a payload executable refuses that
open, the probe answers that the file is in use and that the program holding it
could not be identified - the guard's "could not confirm" dialog, whose Retry
works once the user has closed it - instead of passing the payload as free.
Only the payload executables are tested, not every file the purge removes: a
locked executable proves a program is running and a mapped DLL belongs to a
running executable, while a locked DLL or qt.conf can be the asking process
itself, which is the loop the guard already had. The open is retried once after
300 ms, and only for the files that refused, so a scanner holding a file for a
moment cannot stop an install for nothing.
The Windows probe check now covers that answer: a payload executable held by
the asking process, so that it cannot be named, has to come back as in use
rather than as free.
Co-authored-by: Lukas Brübach <Bruebach.Lukas@bdosecurity.de>
* [Installer] Close running clients over a local socket
The installer could only ask a running client to close with WM_CLOSE, which an instance hidden in the notification area or sitting on its own confirmation cannot answer usefully. Give every client a per-process local socket and have the holder probe ask it there first, falling back to WM_CLOSE for clients from before the socket.
* [Installer] Offer to check again or update by hand when files stay locked
Replace the two-answer guard prompt with force close, close by hand and check again, or cancel. Retry the automatic close while the set of holders changes, and point a cancelled run at the latest release so a user who does not want anything killed has a way forward that is not cancelling the update again.
---------
Co-authored-by: Lukas Brübach <Bruebach.Lukas@bdosecurity.de>
* [Installer] Request quit for hidden Cockatrice instances and record holders in abort telemetry
* [Installer] Fix compiler warnings in holder probe and installer quit handler
* [VisualDeckStorage] Order the includes the way clang-format wants them
* [Installer] Ask a hidden instance to close, then offer to end it once
The previous attempt had the client answer a registered window message, which
cannot work for the update it was reported on: the instance being replaced is a
client from before that message exists, so it never answers, and the user is
left with "Cockatrice is still running" exactly as reported. Every step that can
free the files has to be the installer's, so the client side is reverted.
The probe now posts WM_CLOSE to the instance's own window whether that window is
on screen or only in the notification area: the request does not need a visible
window, an instance in the tray is the one this update has no other way to
reach, and the confirmation Cockatrice answers with is a window Qt shows itself,
so it is painted - unlike one restored from outside. A visible window is brought
forward first, and Qt's bookkeeping windows stay excluded, because a close to the
notification-area one terminates the instance without its confirmation.
A question of the instance's own no longer suppresses the close. Withholding it
protected nobody: the installer asked the user to close a program it had never
asked, and a current client refuses the close while its confirmation is pending
anyway, leaving the question where it is. The launcher mode keeps the
visible-only rule, because a launcher without a window has already answered its
own close and the wait ends it after a grace instead of asking a user who is not
there.
The guard asked the same question until the user gave up: after the automatic
close failed it offered to end the programs, then re-ran the probe and showed
the same dialog word for word when the answer was No, which is how an abort was
reported three times over. It now asks once - Yes ends the Cockatrice, Oracle
and Servatrice programs the payload installs and says what that costs, No
cancels - and what survives an end is named in the abort instead of being put to
the user as the same question again. The uninstaller's guard and the recovery
path after a failed purge follow the same shape.
Kept: the abort telemetry that records the holders, the close result and the end
result as event data.
---------
Co-authored-by: sentry[bot] <39604003+sentry[bot]@users.noreply.github.com>
Co-authored-by: Lukas Brübach <Bruebach.Lukas@bdosecurity.de>
* [Client] Stop reporting local login aborts as missing client features
Local aborts (unusable saved verifier, refused authentication downgrade,
failed scrypt derivation) reuse RespClientUpdateRequired with an empty
feature list, and the handler discarded the reason and advised a client
update. Report the abort with the cause instead, and report a strict
server's challenge_response_auth answer to a credential-less login as the
missing password it is.
* [Client] Keep the login abort messages in one named set
The local login aborts each carried their wording inline at the emit site, so
changing the tone or the remedy meant editing several functions. Collect them
as named constants next to RemoteClient instead, translated at the use sites
with the RemoteClient context so lupdate still picks them up.
* [Client] Reword the login abort explanations to match what happens
The first wording sent users to a server operator to 'migrate' an account,
which operators cannot do, and told them to 'save' a login the client saves by
itself. It also blamed the device for a derivation failure the server's
parameters can cause, and the dialog led with a line that said nothing while
the reason sat in the informative slot.
Say what the client actually does in each case, give the step that works, and
show the reason as the dialog's main text. The unreachable fallback sentence
about the login not reaching the server is gone.
---------
Co-authored-by: Lukas Brübach <Bruebach.Lukas@bdosecurity.de>
* [Client] Report password-reset rate limits instead of a generic failure
sigForgotPasswordError carried no response code, so rate limiting, a
disabled reset feature, an unknown user and a bad token all produced
"contact the server operator". Users who were only rate limited retried,
which kept them rate limited.
* [Client] Move the reset-password warning out of the response switch
---------
Co-authored-by: Lukas Brübach <Bruebach.Lukas@bdosecurity.de>
* [Updater] Refuse a close that arrives while one is already asking
MainWindow::closeEvent() had a re-entrancy guard that returned without touching
the event. A QCloseEvent is accepted unless something ignores it, so a second
close request - an installer replacing our files posts one every 30 s - accepted
the close while the first one was still blocked on its own confirmation. close()
then hid this window, hid the confirmation dialog with it, and left the process
running inside the nested loop that dialog had opened: no window, no way to
answer it, and every later close request accepted into the same dead end. That
is what an update looked like when it could not close a running client, and it
is why the installer could name the process, bring it forward and ask it to
close for minutes without anything happening.
Track how far the shutdown has got instead. While a confirmation is up, a close
from anywhere else is refused, which leaves the question in front of the user;
once the shutdown has been accepted, a duplicate close (Qt delivers the event
twice on some platforms) is accepted again so it cannot cancel the quit.
* [Installer] Leave an instance that is asking a question alone
An instance that answered a close by hiding rather than quitting - the state the
client used to get into - has no visible window left, and the confirmation it is
waiting on is hidden with the window that owns it. The close probe only ever
looked at visible windows, so it restored the main window, posted WM_CLOSE into
the same dead end, and reported "still running" without saying why.
Show every window the instance left hidden before choosing the one to close, and
skip the close when the window it would reach is disabled: a modal dialog of the
instance's own is up, and posting the close then only hides the question again.
Name what is left of the window in the report as well - gone, still open, hidden
again, and whether the thread answers at all - so the log tells a refused close
from one that was never delivered.
* [Installer] Offer to end a Cockatrice that will not close
A close is a request, and an instance that does not answer it - waiting on a
question nobody is looking at, or stuck where a close cannot reach - leaves the
user with no way forward: the files cannot be replaced while it runs, and
clicking Yes again only repeats the request.
After a second close that changed nothing, offer the only thing that can free
the files: end it. The prompt says what it costs - an unsaved deck, a running
game or a setting it has not written yet - the default button is the safe one,
and it is a separate answer rather than a relabelled "Yes", so it is not
reached by clicking through the close prompt. Only cockatrice.exe is ever
ended: Oracle writes the card database in place and Servatrice is a service, so
both are named and left to the user, exactly as they are for the close.
The probe gains an --end-holders mode for it, and the uninstaller's guard gets
the same offer.
* [Installer] Never name the process that is doing the asking
The old uninstaller is started with _?= so it runs in place, from the very
directory it is emptying, and an installer can be downloaded into the folder it
installs to. Both are then processes running from the directory, so the probe
named them - and the guard asked the user to close the program that was asking.
That question cannot be answered: a close and the end offer both skip anything
that is not cockatrice.exe, so the guard repeated itself until the update or the
uninstall was cancelled.
The probe and the processes that started it are now never named, by either
lookup. The walk up the parent chain stops at the first program the payload
installs, because in an in-app update the installer's own parent is the
Cockatrice instance being replaced - the one the guard exists to name.
The Windows check held its locked file in the script that starts the probe, so
that holder is excluded by design now and the check read it as "nothing found".
It holds the file from a process of its own instead, and asserts the exclusion
itself: a file that only the asking process holds comes back as nothing.
* [Installer] Restore only the windows a user could answer
Showing every hidden window an instance owns put Qt's own bookkeeping windows on
screen: the tray icon keeps one to receive the TaskbarCreated message, styled
WS_OVERLAPPED and titled QTrayIconMessageWindow. Restored, it is an empty white
rectangle in the taskbar with no title bar to close it - it has no window
procedure that would hide it again, and the process is busy in the very loop the
restore is meant to break, so it stayed until the process was force-ended.
Only framed, non-empty windows are restored and considered for the close now.
That also keeps the close off Qt's tray window, which answers WM_CLOSE by asking
the application to terminate - a quit that skips the confirmation the close is
supposed to reach.
* [Installer] Uninstall the previous version before installing over it
A regular install over an existing one asked "a previous version must be
uninstalled", then ran the old uninstaller interactively. That uninstaller
put its confirmation page behind the installer window, where it could not be
answered, and it only removed the installation after the installer had
already written the new one over it.
Run the previous version's uninstaller silently instead. The installer has
already asked the question the user answered, so a second, hidden confirmation
is pure risk: with /S there is no window to be behind, and ExecWait is then
waiting for something that cannot outlive it. QuietUninstallString is used
when the old installation has one, otherwise /S is appended.
ExecWait only waits while the uninstaller runs in place: one that did not get
a usable _?= copies itself to %TEMP% and returns at once, while the copy keeps
deleting - which is how an uninstall can finish after the install it was
supposed to precede. Log the exit code and wait for the old cockatrice.exe to
disappear before restoring user data or writing anything, and abort with a
message instead of installing into a directory that is still being emptied.
Cancelling the "must be uninstalled" prompt now quits the installer; it used
to skip the page and install over the old version anyway.
* [Installer] Do not show windows the application has hidden
The guard restored every hidden window an instance owned before asking it to
close, so that a confirmation hidden together with its window would come back.
Windows does as it is told and shows them; the application does not, because Qt
still believes they are hidden and never paints them. What the user gets is a
blank white rectangle, on screen and on the taskbar, that answers nothing - and
the question it was meant to bring back is not answerable in it either.
An instance with no window a user could answer is named and left to the user
instead, and the prompt says what to do about it: open it from the notification
area. Ending it stays the last resort, offered once a close has been asked for
twice.
The abort after a failed uninstall now says what it saw - the old version is
still there after its uninstaller ran - instead of promising a removal that is
not happening, which is the state a cancelled old uninstaller leaves behind.
---------
Co-authored-by: Lukas Brübach <Bruebach.Lukas@bdosecurity.de>
* [Translation] Name languages without reading the translation files
The language selector derived every label from the language's own .qm file:
it loaded the file and translated the literal "English" in the "i18n"
context. lupdate could never extract that literal from the callers, which
pass the DEFAULT_LANG_NAME macro, so a dummy QT_TRANSLATE_NOOP in
cache_settings.cpp existed to give it something to extract. #7050 deleted
that dummy, the next lupdate push dropped the string from Transifex, and
the following pull rewrote the .ts files without it. lrelease only emits
translated messages, so 22 of 25 .qm files returned an empty name.
The names are data, not translations: a selector has to be readable before
any translation is loaded, and a label must not depend on the translation
of the language it names. TranslationLoader::languageDisplayName now holds
them next to availableLanguages, falling back to Qt's native language name
and then to the code, so a language added later still gets a label.
Also select the configured language by item data instead of by matching the
translated "English" string, which pointed at the wrong row whenever the
lookup came back empty, and restore that English entry's own name: en_US is
"English (United States)" so it can no longer be confused with the
untranslated default.
Oracle shared the code path, so its wizard language page is fixed too.
* [Translation] Name the default language selection once
Review follow-up: log the unresolvable language through the page's own
category, and move the en_US entry both selectors fall back to for an unset
setting into TranslationLoader::defaultLanguageCode.
---------
Co-authored-by: Lukas Brübach <Bruebach.Lukas@bdosecurity.de>
* Retranslate all open tabs on language change
* Address review: deeper Visual DB retranslation, index-based Archidekt size logic
In Qt 6, setItemText() on the current item emits currentTextChanged but not
currentIndexChanged, so connecting the Archidekt combos to currentTextChanged
fired a search on every language change. Both combos now use currentIndexChanged,
and buildSearchUrl() compares the size-logic combo by index so translated labels
cannot drop sizeLogic.
* Derive display mode button text in one place
* [Models] Use QCollator for natural, cached sorting
QString::localeAwareCompare rebuilds its collation data for every
comparison, and it compares numbers as text. Keep one QCollator per model,
enable numeric mode, and use it for the card name, set and user name sorts.
Add a test that sorts "Forest 1", "Forest 2", "Forest 10" through
DeckListSortFilterProxyModel and checks the natural order.
* [Models] Sort naturally under the C locale too
Numeric mode only takes effect through a real locale: under the C locale,
which headless CI processes run in, QCollator silently falls back to string
order and sorts "Forest 10" before "Forest 2". That made collator_sort_test
fail on every Linux leg while passing locally.
Add makeNaturalSortCollator(), which enables numeric mode and pins English
only when the process locale is the degenerate one, and build the models'
collators through it. Users with a real locale keep their own collation.
* [Models] Fix compile of natural sorting (QStringView mismatch)
* nit: initialize collators in member initializer lists
* nit: namespace makeNaturalSortCollator under NaturalSort
---------
Co-authored-by: Lukas Brübach <Bruebach.Lukas@bdosecurity.de>
The local picture loader walked the whole custom picture tree every ten
seconds on the GUI thread. Watch the root and every subdirectory with
QFileSystemWatcher, debounce bursts of events into a single 500 ms refresh,
and keep a five-minute fallback poll for filesystems where watching fails.
Add an explicit-path constructor and indexedPaths() so the indexing can be
tested without SettingsCache, and emit indexChanged() after a rebuild.
Verification: cockatrice built.
Co-authored-by: Lukas Brübach <Bruebach.Lukas@bdosecurity.de>
* [Network] Add a configured QNetworkAccessManager factory
Every HTTP caller built its own QNetworkAccessManager, so the system proxy
and a transfer timeout were easy to omit. Add NetworkManagerFactory::create()
as the single place that applies the process-wide proxy configuration and a
transfer timeout, and use it at all 14 construction sites.
Per-site settings such as disk caches and manual redirect policies stay with
their callers.
* [Tests] Link libcockatrice_network into the tests that compile its headers
release_channel.cpp and card_picture_loader_worker.cpp now include
libcockatrice/network/network_manager.h, but the test targets that compile
them only linked libcockatrice_settings, so the include directory never
reached them and every platform failed with "file not found":
loader_local_matching_test and picture_loader_benchmark_test.
libcockatrice_network is an INTERFACE library, so linking it costs these
targets nothing but hands over the include path.
---------
Co-authored-by: Lukas Brübach <Bruebach.Lukas@bdosecurity.de>
Read the avatar with setAutoTransform(true) so camera orientation is
applied, and ask QImageReader for a scaled decode at MAX_AVATAR_DIMENSION
instead of decoding a full-resolution photo and scaling it afterwards.
Verification: cockatrice built.
Co-authored-by: Lukas Brübach <Bruebach.Lukas@bdosecurity.de>
* [QtUtils] Add parent lookup, layout clearing, scroll area and monospace helpers
* [QtUtils] Adopt shared Qt helpers across the client
* [QtUtils] Add layout creation and relayout helpers
* Swap parameter order (spacing before margin) and remove static from relayout()
* [QtUtils] Use relayout() after rebuilding dynamic layouts
* [QtUtils] Use makeLayout() for zero-margin layouts
* Consolidate setSpacing() calls into makeLayout() using new parameter order
---------
Co-authored-by: Lukas Brübach <Bruebach.Lukas@bdosecurity.de>
A palette preview repaints widgets through applyStyleAndPalette without running
themeChangedSlot, so generator-owned pixmaps stayed at the previous palette
until the next full theme switch. Clear them where the palette is applied.
Co-authored-by: Lukas Brübach <Bruebach.Lukas@bdosecurity.de>
* Retranslate Home tab when the language changes
TabSupervisor::retranslateUi() never included the Home tab, and
HomeWidget set its button text only once, when the buttons were
created. Switching languages left the Home tab in the old language
until it was reopened.
Store the Home buttons as members, add HomeWidget::retranslateUi()
to set their text, and call it from TabHome and TabSupervisor. The
connect button text still only changes for the Connecting,
Disconnected and LoggedIn statuses, as before.
* Address review: plain ClientStatus, drop optional and placeholder
* Refactor display type toggler and add language change handling in VisualDeckDisplayOptionsWidget
* Address review: extract updateDisplayTypeButton(), document changeEvent
* Restore trailing newline in visual_deck_display_options_widget.cpp
* [CardDB] Avoid copying printing maps and property strings
Reading a card's sets was a measured hotspot: `SetToPrintingsMap setMap =
card->getSets();` copies the whole QMap, and then iterating it with a
non-const iterator detaches and deep-copies every PrintingInfo, including
its property hash, on first use. A micro-benchmark of that exact pattern
(copy plus non-const iteration) came out about 19x slower than a const
reference over the same data.
The querier, the cache writer, the XML writer, the picture loader and the
printing/set selection widgets now take `getSets()` by const reference and
only use const iterators. The tests that iterate generated data are
covered by the existing card database and picture loader suites.
String accessors were also returning by value where the value lives in a
hash that has already been materialized:
- CardInfo::getProperty() and the legacy getCardType()/getColors()/
getManaCost()/getCmc()/getPowTough()/getLoyalty() accessors,
- PrintingInfo::getProperty(), getUuid(), getFlavorName(), getArtist(),
and getSet(),
- ExactCard::getPrinting() and getCardPtr(),
- CardSet::getShortName(), getLongName() and getSetType().
getPrinting() in particular copied a PrintingInfo on every call, and a
copy of an unmaterialized PrintingInfo has to take the properties mutex.
Callers that need an owned copy still get one; the rest just read.
Other allocation removals in the same path:
- LazyPropertiesHash::ensureMaterialized() checks an atomic flag before
taking the mutex, so property reads after the first one do not lock.
The flag is release-stored after materialization and acquire-loaded by
readers/copies, preserving the existing locking for the first read.
- CardSet::getCorrectedShortName() no longer rebuilds its QSet of
reserved Windows file names on every call.
- CardSet::getIsKnownIgnored() no longer formats the release date into a
throwaway string just to check its length.
- The querier builds its card-type/format count caches without
`cards.values()` or a properties key list, reserves its result lists,
and picks a random card without materializing `cards.keys()`.
- The querier's pinned-printing path reuses the CardInfoPtr it already
has instead of looking the card up again by name.
- CardDatabaseCache/XML serialization copy the alt-name QSet by
reference.
- The picture loader compiles its !prop!/!set! regular expressions once
instead of per URL template, and walks the transform map with an
iterator instead of keys() plus operator[].
Full build and ctest (628 tests) pass.
* [CardDB] Address review comments
PrintingInfo is a value type: it is stored in QLists and returned by value
from the querier, so a const-ref getter can hand out a reference into a
temporary. Return the QStrings and the CardSetPtr by value again.
The constFind lookup stays, so a property read still never materializes the
lazy property hash - only the last shared-data refcount is paid.
* [CardDB] Undo the const-ref QString return changes
QString is implicitly shared, so returning by value costs only a refcount
and avoids handing out references into a lazily materialized hash. The
constFind lookups stay, so a property read still never materializes the
hash.
* [CardDB] Use QMap::asKeyValueRange() in transformUrl
The range is a view over the map's own iterators and yields references to
the stored key and value, so it keeps the no-allocation property of the
manual const iteration.
---------
Co-authored-by: Lukas Brübach <Bruebach.Lukas@bdosecurity.de>
* [VisualDatabase] Use scoped QSignalBlockers for filter updates
updateFilterModel() blocked signals on the filter model and its filter tree,\nthen unblocked each of them twice. The second pair was a no-op only because\nthe first pair had already restored the default. Use two scoped\nQSignalBlockers instead, so the models are unblocked exactly once before the\nexplicit changed()/layoutChanged() notifications.\n\nVerification: cockatrice built.
* Drop stale typo note from QSignalBlocker comment
---------
Co-authored-by: Lukas Brübach <Bruebach.Lukas@bdosecurity.de>
* [CrashReporting] Breakpad everywhere, session detection, symbol CI (inc 3)
Breakpad on all platforms (no crashpad packaging); remove the Windows
MiniDumpWithFullMemory filter and the installer WER LocalDumps full-dump
registration. QLockFile per-process session tracking replaces the
Windows-only .dmp scan, placed after parser.process and the instance
hand-off so help/version exits and handed-off secondaries never leak a
lock. CI: Release-only DSN bake from SENTRY_DSN secret, tag-only
sentry-cli symbol upload (org cockatrice, project native).
Tests: 533/533 ctest green, format clean, code gate PASS.
* [CrashReporting] Enable http_retry, harden symbol upload, fix NSIS context
Sets sentry_options_set_http_retry so the cache_keep-held crash
envelope in Ask mode is written with a retry filename and actually
flushes on giveConsent() instead of sitting in <db>/cache until
pruning.
Adds continue-on-error to both sentry-cli install steps so a Sentry
registry outage or broken Git Bash install cannot block release
assets.
Cleans the legacy WER CrashDumps folder under SetShellVarContext
current (under all it resolves to ProgramData, not the user's
AppData) in installer and uninstaller; registry delete works in
either context.
* [CrashReporting] Extract the startup crash-report flow from main() per review
* [CrashReporting] Break up the CrashReporter class doc per review
* [CrashReporting] Drop the invalid --release from the symbol upload
`debug-files upload` never accepted `--release` (checked against the pinned
sentry-cli 3.8.0: the subcommand declares --org/--project and a set of
type/id/scan options, and execute() reads no release; the only global flags are
--url/--header/--auth-token/--log-level/--quiet/--allow-failure). Clap rejects
the unknown flag before any upload happens, so every tag build has been shipping
without symbols, with continue-on-error keeping the release pipeline green.
Release association is also the wrong mechanism: debug information files are
matched to events by debug ID via the event's debug_meta. The Sentry release of
a native event comes from the event itself (COCKATRICE_SENTRY_RELEASE), not from
the DIF upload. Remove the flag and the now-unused SENTRY_RELEASE env from both
invocations (Linux, and Windows/macOS).
The symbol upload stays continue-on-error (a Sentry outage must not block the
release assets); revisit once it has proven itself on a tag.
* [CrashReporting] Split debug information out of packaged release binaries
-g is required so the CI symbol upload has DWARF to send, but it was left in
the shipped binaries too: the Linux Release client is 563 MiB, 552 MiB of it
debug info, and every distribution channel pays for that.
Keep the build tree unchanged (CI uploads symbols from build/) and strip only
the packaged copy. A CPack pre-build script extracts each ELF's debug info to
COCKATRICE_DEBUG_SYMBOL_DIR and then runs `strip --strip-debug --strip-unneeded`.
.eh_frame is allocatable and survives, so unwinding is unaffected, and the build
ID is identical in the stripped binary and the sidecar.
Measured on a local Release build (GCC -O2 -g, WITH_CRASH_REPORTING=ON),
bytes:
binary build tree packaged before packaged after .debug sidecar
cockatrice 590,722,560 590,722,560 28,086,480 579,623,576
oracle 53,432,608 53,432,608 1,649,888 52,075,344
tar.gz of both trees: 289,964,521 -> 19,018,058 (~15x smaller).
Only ELF (Linux/BSD) is handled: Windows already links /DEBUG with separate
PDBs, and macOS extracts a dSYM from the bundle. Opt-out via
COCKATRICE_SPLIT_DEBUG_SYMBOLS; skipped with a warning when objcopy/strip are
missing. CPACK_PRE_BUILD_SCRIPTS (not CPACK_INSTALL_SCRIPTS, which runs before
staging) invokes it, and the script refuses to run unless CPACK_TEMPORARY_DIRECTORY
is a real CPack staging tree so a missing variable can never turn its glob into
a filesystem-wide walk.
* [CrashReporting] Upload Windows PDBs from where Ninja actually puts them
The Windows branch of the symbol upload looked in build/<target>/Release, which
is the Visual Studio generator's layout. The Windows leg uses Ninja, which is
single-config: linker PDBs land in build/<target>/, next to the binaries — the
same paths the "[Windows] Upload PDBs" artifact step below globs with
if-no-files-found: error.
So paths was always empty on Windows, the step printed "No symbol paths found,
skipping symbol upload" and exited 0: a silent no-op behind continue-on-error.
Every Windows build would then report "A required debug information file was
missing" for cockatrice.exe with a PDB-style debug id (GUID-age), exactly as if
no upload step existed at all.
Use the artifact step's globs. Also turn the empty-path case into a failure in
both symbol steps: a silent skip is indistinguishable from a successful upload
(which is how the wrong path went unnoticed), and continue-on-error already
keeps a broken upload from failing the release.
Verified by extracting both steps' run blocks and executing the Windows branch
against a Ninja-style tree: it now passes the three linker PDBs to sentry-cli
(and not the CMakeFiles/*.pdb intermediates), and exits non-zero with an
::error:: annotation when there is nothing to upload.
* [CrashReporting] Archive PDBs on tag builds so a failed upload is recoverable
The "[Windows] Upload PDBs" artifact step skipped tag builds
(github.ref_type != 'tag'), so for a release the symbols existed only inside
the Sentry upload — which is continue-on-error. A failed upload therefore lost
the ability to symbolicate that release permanently: no rebuild produces the
same PDB, since the debug id is a per-build GUID+age. That is how every
Windows crash reported so far became unreadable.
Archive the PDBs for tag builds as well, and order the step after
"Upload to release" so a missing PDB can never block a release from being
published. Retention stays at the repository default; nothing consumes the
artifact automatically, it is the manual recovery path
(gh run download <run> -n <name>-PDBs, then sentry-cli debug-files upload).
Verified by re-parsing the workflow and asserting the step order
(release -> PDBs -> attest) and the broadened condition.
* [CrashReporting] Upload the Windows executables with their PDBs
Sentry needs two things per image on Windows: symbolication and stack
unwinding. The PDB covers symbolication, but for a 64-bit PE the Call Frame
Information (.pdata/.xdata) lives in the executable and is not copied into the
PDB, so uploading PDBs alone leaves the image at "Stack Unwinding: Missing".
That is exactly what the event's Images panel showed for cockatrice.exe, while
Sentry's own Microsoft symbol server resolved KERNELBASE.dll and ucrtbase.dll
from the same dump.
sentry-cli also needs the pair in the same invocation: only then does it
associate the PE's code identifier with the PDB's debug identifier, otherwise
"the identifiers diverge and Sentry might not be able to resolve the correct
file for symbolication" (docs: Debug Identifiers).
Do the same for macOS, where the dSYM carries the debug information but the
Mach-O carries the unwind information, and include the executables in the
symbol artifact so a re-upload still has the pair to offer.
Verified by extracting the step and running both branches against stub
harnesses: Windows passes the three exe/pdb pairs in a single sentry-cli
invocation and still fails loudly on an empty tree; macOS passes the dSYM and
the Mach-O together.
* [CrashReporting] Upload the shipped Qt runtime DLLs too
Qt ships no PDBs with its MSVC SDK, so every Qt frame in a crash report
resolved to nothing: an image with neither PDB nor PE has no unwind
information (on 64-bit Windows the Call Frame Information lives in the PE
and is not copied into the PDB) and no symbol source to fall back on.
The list is required-runtime.txt, the manifest .ci/compile.sh already
asserts the installer against, so what gets uploaded is exactly what an
install carries. Qt frames now get unwinding plus export-table names; line
numbers still need Qt's debug information files, which the SDK omits.
The guards are not cosmetic: the step runs under `bash -eo pipefail`, where
a sed over a missing file or an unmatched ls glob aborts the whole step, and
continue-on-error makes that look identical to a successful upload.
* [CrashReporting] Upload only the binaries, archive symbols only on tags
The Linux leg handed sentry-cli whole build directories: 471 candidate files,
of which 7 were object files that can never match a stack frame. Only the
three linked binaries carry the build-ids Sentry looks up; the packaged copy
is split and stripped but keeps that same build-id.
The Windows symbol archive ran on every packaging build even though only tag
builds upload to Sentry, so a PR build paid for an artifact nothing could
ever consume. Tags keep it as the durable copy for a failed upload.
---------
Co-authored-by: Lukas Brübach <Bruebach.Lukas@bdosecurity.de>
