Now that the client is pushing all changes of state, we don't need
to track the requested URLs anymore and risk that the way that we
reseted that list could leave a few entries in Finder's cache outdated.
We can remove the same code from other platforms in a later release,
a bit earlier than a week before beta1.
This fixes a few issues with the new FinderSync integration on OSX which
can't easily clear its status cache when receiving an UPDATE_VIEW message
except by unregistering the folder, but which causes flickering.
This fix should also make the regeneration of the cache unnecessary on other
platforms through possibly expensive RETRIEVE_FILE_STATUS commands.
To allow forwarding all OK states of sync jobs to the socket API
we need to make sure that the vector doesn't contain unneeded items.
This initially was intended to force updating the metadata for parent
folders, but since then the should_update_metadata flag check was
added both here and in PropagateDirectory::finalize where the metadata
is actually updated for parent folders. We can safely remove the inclusion
of all remote directory items.
The signal is emitted in this case if the instruction is NONE
but in the only ultimately connected slot to this signal,
SocketApi::slotSyncItemDiscovered, we return early according
to the same condition.
The emission of the same signal at the end of treewalkFile remains
and take care of the normal cases.
This makes sure that users won't have to go into settings
to activate the extension. This might upset users by forcing
them to re-disable it on each oC upgrade, but if this becomes
an issue we should then make this configurable in the client's
settings itself.
This allows developers to build and run the extension by default.
Official packages bundles will be re-signed after the build, we
The SocketApi prefix can be set at configure time through cmake and
should match the key that will be used to sign the whole .app bundle
(including the embedded FindexSync .appex bundle).
Since we use that class to lookup the NSBundle using bundleForClass
use a more specific name to avoid any clash with any othe liferay
extension. I couldn't figure out from the documentation if that is
only resolved using the class name, but found some warnings on
stackoverflow and better be safe than sorry for what it costs.
Since we periodically recreatet the QNAM owned by the Account,
the web view could end up holding a dangling pointer to it.
Let the web view use its own QNAM, but make sure that the cookie
jars are shared so that we can grab the session token given
by the server.
It was only used on OS X and couldn't be used by the FinderSync
extension since that one runs in a sandbox. So use the same system
to load images in the legacy extension by shipping them in the
extension bundle instead of the owncloud.app bundle.
This is also given that the legacy extension needs padded icons
while the FinderSync one needs unpadded icons.
Since GIT_SHA1 would need to be updated in config.h, all files
including it would be rebuilt by make.
Reduce the number of files to rebuild by moving this variable
to version.h instead.
The archive buildaction causes this. Use the default build while
forcing the Release configuration instead.
In both cases the result will end up in SYMROOT.
This uses the new official API to show overlay icons and add our
custom context menu entry instead of hooking directly into the
Finder process and intercept drawind routines.
A dummy desktopclient target is also in the project to allow debugging
directly in Xcode while the official client can be started from the
command line. Otherwise Xcode won't allow attaching to the debugee.
Dummy icon files have been added while we get proper icon produced.
We can't use the old icons since what we use for the legacy shell
integration is already padded according to where the badge should
appear on the full icon.
This prepares the switch to the official FinderSync API on Yosemite
which requires the extension to run in a sandbox. This complicates
the usage of a local socket to communicate with a non-sandboxed GUI
client. An NSConnection is easier to use in this case, which we can
use as long as the server name (i.e. Mach port registered name) is
prefixed with the code signing Team Identifier.
A placeholder server implementation is also added to the client's
SocketApi which basically reproduces the interface of a QLocalSocket.
Most of the references to individual sockets we're only using
QIODevice methods so the type was simply reduced. A typedef to
replace the QLocalServer was the only other part needed.
Since our DLLs version number aren't bumped, a same file size could
cause a DLL not to be installed to replace an old one.
This eventually can cause a crash of explorer if an incompatible version
of OCUtil.dll is installed alone.
Fix the issue by defining LIBRARY_IGNORE_VERSION. This also makes sure
that installing an old version of the client over a more recent one
if we decide to add version numbers to those DLLs.
Do not request the status of all entries in the cache. Instead force Finder
to request the ones that it deems necessary by keeping the old statuses
in a separate dictionary which are only used while the new status arrives.
removeIcons isn't called and clearFileNameCacheForPath was always
called with a "nil" path.
Remove the return value of askForIcon which was always 0, and use
that value explicitly at the only call site.
Remove the "-1" code path in iconByPath since setIcons prevents
-1 from getting into _fileNamesCache in all cases.
Since QNonContiguousByteDeviceThreadForwardImpl::reset will
call UploadDevice::reset with a BlockingQueuedConnection, this
allows us to reset the HTTP channel along with its buffers
before they get the chance to be reused with a subsequent request.
Do not request the status of all entries in the cache. Instead force
explorer to request the ones that it deems necessary by keeping the old
statuses in a separate dictionary which are only used while the new status
arrives.
Use a QSharedPointer to keep the same ownership and
continue passing the SyncFileItems as a const& when
ownership isn't taken. This allows sharing the same
allocations between the jobs and the result vectors.
This saves about 20MB of memory (off 120MB) once all
jobs are created.
Since the QLocalServer parent of the QLocalSockets gets destroyed
after the _listeners QList, onLostConnection might try to update
an already destroyed list.
Fix the issue by simply making sure that _localServer is destroyed
first.
In some cryptic cases where the getetag property wasn't returned by
the server, we might be trying to c_strdup a null pointer in
csync_vio_file_stat_copy.
At least avoid crashing in this case by looking for
CSYNC_VIO_FILE_STAT_FIELDS_ETAG, like csync_vio_file_stat_destroy
does.
The default constructor of the iterator points to NULL, which makes
it != end() but invalid to dereference.
Use an integer index instead to keep 0 as a valid default value that
can always correctly be checked against size().
Also make sure that no data is shared between threads by making the
csync_vio_file_stat_t copyable and passing it as const.
Deleting the QLocalSocket while iterating the QList with qDeleteAll
would trigger onLostConnection, which would modify the list mid-air
and leave dangling pointers in it.
Make sure that we define NDEBUG in all configurations.
Also remove inconsistent defines:
WIN32: We don't use
_DEBUG;_WINDOWS;_USRDLL: Should be defined by the compiler if necessary
OCCONTEXTMENU_EXPORTS: We currently don't use dllimport anyway
This also update the binary submodule with an updated build.
Since each new connection to the socket API would trigger a broadcast
of REGISTER_PATH to all existing connections, opening the context menu
would trigger a SHChangeNotify call of the root directory through
the overlay icon extension, which is currently also connected to the
socket API, waiting for changes.
Fix the issue by sending the initial REGISTER_PATH automatic response
only to the connecting socket.
This moves the crash handler installation during the OCC::Application
contruction. This still leaves a window where crashes wouldn't be
caught, leaving the QtSingleApplication and theme initialization
code unreported, but isn't requiring any refactoring for now.
Issue #2952
The config currently only disable the reporter, but still leaves
the crash handler installed on startup. This causes conflicts
when running those builds in lldb.
Work around the issue by requiring the crash reporter to be
disabled in the settings before owncloud can be run in lldb.
This has the negative side-effect of not covering the
initial ConfigFile loading with the crash handler.
Issue #2946
This attempts to replicate what i686-w64-mingw32-gen_sym_files
is doing for us with MinGW by parsing the output of otool -L
to get the library name actually used by the executable.
Import a fix from libcrashreporter-qt that should remove
usage of symbols not available in XP's msvcrt.dll
This should help solving #2907 with the crash reporter enabled.
sendMessage would only be called if there were options to
be sent to the running application.
Fix the issue by having explicitly named messages and always
send the show settings message.
Issue #2374
There is a race condition that can initialize the RemotePathChecker
instance concurrently on the same address and cause a crash when locking
the mutex. The reason is that local static initialization is not
thread-safe with MSVC2013.
Fix the issue by using call_once to initialize a static unique_ptr
instead.
This could be related to some reports of issue #2836.
CMake is using "c++" as the compiler and the clang detection fails.
Since we require C++11 at this point, enable it in all cases.
Removed the MSVC todos since we don't plan supporting it yet.
50ms is sometimes not enough when the client is busy synchronizing.
Wait up to 500ms for the client to answer before we give up and show an
empty menu title.
Ideally we should request the title before the watched directory list, but
the list is currently sent implicitly on connect.
The context menu will now show "Share with ownCloud" instead of "Share" as
it does on other platforms.
This also updates the submodule to point to matching binaries.
It seems like verclsid.exe hangs on our class IDs when invoked
through explorer.exe for 5-10 seconds. It doesn't hang if I
invoke the same command line from cmd.exe, so there could be some
process parameters that don't play well with our extra thread
or to the pipe connection that we do in it.
Delay creating the RemotePathChecker thread until the first
IsMemberOf call. verclsid.exe only seems to instantiate a object
of each registered class, without actually using them, so we
can use this as a workaround.
This should be fixing issue #2680.
Windows finds DLLs using PATH or the directory of the process'
executable. By outputing those dependend DLLs together with
owncloud.exe, the developer only need to have OpenSSL's bin
and the qtkeychain build directory in his PATH to let the
dynamic linker find them.
As the documentation of RUNTIME_OUTPUT_DIRECTORY points out,
this only affects windows as other platforms don't consider
libraries as runtime targets.
- qtkeychain isn't necessarily in a qt5keychain subdirectory.
- Explicitly add OpenSSL to the include directories
- Make sure that the USE_OUR_OWN_SQLITE3 code is initialized
for csync by calling find_package.
Jocelyn Turcotte