By default QNetworkReply::errorString() often produces messages like
"Error downloading <url> - server replied: <reason>"
but the "downloading" part invariably confuses people since the
error might very well have been produced by a PUT request.
This commit produces clearer error messages for HTTP errors.
Additionally:
* Remove some unnecessary null checks from slots connected to
network job signals and document that these signals never send
null replies.
* There was a bug where AbstractNetworkJob::_timedout wasn't
set when derived classes overrode slotTimeout. We now ensure
it's always set by disallowing overrides of slotTimeout.
Instead it now calls onTimedOut, which allows custom handling.
* Several subclasses declared errorString, isTimedOut. Move
these to AbstractNetworkJob.
* Unify handling of OC-ErrorString (via the new, general
Job::errorString)
* Add documentation in various places.
* make target duration a client option instead of a capability
* simplify algorithm for determining chunk size significantly
* preserve chunk size for the whole propagation, not just per upload
* move options to SyncOptions to avoid depending on ConfigFile
in the propagator
* move chunk-size adjustment to after a chunk finishes, not when
a new chunk starts
It is possible to create files with filenames that differ
only by case in NTFS, but most operations such as stat and
open only target one of these by default.
When that happens, we want to avoid uploading incorrect data
and give up on the file.
Typically this situation should never occurr during normal use
of Windows. It can happen, however, when a NTFS partition is
mounted in another OS.
* For requests:
- reuse the original QNetworkRequest, so headers and attributes
are the same as in the original request
- determine the original http method from the reply and the request
attributes
- keep the original request body around such that it can be sent
again in case the request is redirected
* Simplify the interface that is used for creating new requests in
AbstractNetworkJob.
Previously this wasn't happening for errors that were not
NormalErrors because they don't end up in the blacklist.
This revises the resetting logic to be independent of the
error blacklist and make use of UploadInfo::errorCount
instead.
412 errors should reset chunked uploads because they might be
indicative of a checksum error.
Additionally, server bugs might require that additional
errors cause an upload reset. To allow that, a new capability
is added that can be used to advise the client about this.
Use a QMap to avoid using a full hashtable for only a few entries, and
clear the QMap once we're done with the measuring. This saves a few
hundred bytes per job during propagation that would otherwise only be
freed at the end of the sync.
Same fix as in commit 60c101d9
From the crash reporter:
Crash
EXCEPTION_ACCESS_VIOLATION_READ at 0x4
qnetworkreply.cpp in QNetworkReply::request at line 476
propagateupload.cpp in OCC::PUTFileJob::slotTimeout at line 100
moc_abstractnetworkjob.cpp in OCC::AbstractNetworkJob::qt_static_metacall at line 98
qobject.cpp in QMetaObject::activate at line 3716
moc_qtimer.cpp in QTimer::timeout at line 192
qtimer.cpp in QTimer::timerEvent at line 247
qobject.cpp in QObject::event at line 1267
qapplication.cpp in QApplicationPrivate::notify_helper at line 3722
qapplication.cpp in QApplication::notify at line 3505
qcoreapplication.cpp in QCoreApplication::notifyInternal at line 932
This fixes an issue in which too many jobs are started un parallel
while uploading many files, which could cause too much memory usage as the
chunks are stored in memory.
Probably the fix for #4611
* Add checksums/supportedTypes and checksums/preferredUploadType
capabilities. The default is that no checksum types are supported.
* Remove the transmissionChecksum config option. Servers must now
use the capabilities to indicate that they are fine with the
client sending checksums.
Note: This intentionally breaks brandings that overrode
Theme::transmissionChecksum. The override must be removed and the
server's capabilities must be adjusted to include the new values.
Helps with small file sync #331
When I benchmarked this, it went up to 6 parallelism and
was about 1/3 faster than the previous fixed 3 parallelism.
Doing more than 6 is dangerous because QNAM limits to 6 TCP
connections and also the server might become a bottleneck.
Should also help for #4081
Added a new "chunkSize" entry in the General group of the owncloud.cfg
which can be set to the size, in bytes, of the chunks.
This allow user with hude bandwidth to select more optimal chunk size
Issue #4354
Server older than 8.1 cannot cope with invalid char in the filename
so we must not send them from the client. We were already checking
for new files, but not for renames or new directories.
https://github.com/owncloud/enterprise/issues/1009
* Ensure every time a file becomes a directory or the other way around
the item is flagged as INSTRUCTION_TYPE_CHANGE.
* Delete the badly-typed entity if necessary in the propagation jobs.
* Compute the content checksum (in addition to the optional
transmission checksum) during upload (.eml files only)
* Add hook to compute and compare the checksum in csync_update
* Add content checksum to database, remove transmission checksum
We changed the discovery code not to ignore files whose filename contains
charachter invalid on windows. (Because newer versions of the server
supports them)
Servers older than 8.1 will just say "Bad Request" as an error and it's a
regression against previous client version. So keep nice error even with
older server.
Relates to #3736
Because that's what's going on. A job can 'complete an item' or 'finish'.
Note that several jobs could complete the same item: a new directory
will complete on the PropagateRemoteMkdir and the PropagateDirectory
jobs.
Previously, PropagateDirectory jobs didn't emit the completed() signal.
Now that they do, we need to make sure to not add extra lines to the
protocol widget for them.
To accomplish that, the jobCompleted() signal now also contains the job
that completed the item.
In particular the 'unsupported client version' error message
is now visible to the user when trying to connect to a
server that no longer supports the current client version.
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.
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.
This is largely a guess, but this is the only place where we use
a QIODevice to push data through QNAM and that the QIODevice isn't
a direct child of the QNetworkReply.
Fix the issue by making sure that we don't go back to the event loop
and possibly handle network events between the destruction of the
upload QIODevice and the QNetworkReply, which might lead to QNAM
dereferencing a dangling QIODevice pointer.
QFileInfo has to be refreshed if the underlying file has been
modified in between. That is dangerous so ckamm and me decided
to eliminate the QFileInfo based implementations.
This was triggered by a bug that the client uploaded files that
it should not have.
This crash was uncovered when testing parallel chunks when one chunk
finished the upload. The propagate was then finished with the other
UploadDevices still being in-flight in the QNAM.
This should solve #2675 and #1981
By preloading the chunks in memory before sending them, we don't keep the
file open and therefore we let other program open the file for writing.
If the file is modified between two chunks, we detect that and abort anyway
The server is slow to process each chunk. As a result the time between
the chunks is not neglectible. Therefore we gain a lot of time by
uploading the chunks in parallel. So the bandwitdh is used when the
server is waiting between chunks.
* Use a shared pointer to Account everywhere to ensure
the instance stays alive long enough for a sync to terminate
* Folder is now tied to an AccountState
* SyncEngine and OwncloudPropagator tie to an Account and use that
for all jobs they run
Issue: Since the setup wizard currently always replaces the
account, it will always wipe all folder definitions, even when
the actual changes to the account were minor.
Christian Kamm