This commit is the 1/3 part of the fix for the following security
vulnerability:
- CVE-2021-42076 DoS via excess length messages
The issue has been reported by Matthias Gerstner <mgerstner@suse.de>.
(cherry picked from commit e33c81b835)
When disconnected a NOTE log entry is made in the log, when connected it just prints which seems inconsistent to me:
```
[2020-06-01T11:50:55] INFO: AES256-GCM-SHA384 TLSv1.2 Kx=RSA Au=RSA Enc=AESGCM(256) Mac=AEAD
/home/jonathan/barrier/src/lib/net/SecureSocket.cpp,872
connected to server
[2020-06-01T11:52:43] INFO: entering screen
```
vs.
```
[2020-06-01T11:59:01] NOTE: disconnected from server
/home/jonathan/barrier/src/lib/barrier/ClientApp.cpp,315
```
Since everything else logs with a timestamp I suggest changing this to NOTE as well for connection events being logged.
Fixes high cpu usage spikes on win10.
When queue was containing messages of only non-QS_POSTMESSAGE type the
"while (m_buffer->isEmpty())" busy-looped in EventQueue::getEvent
since isEmpty was true (checked only QS_POSTMESSAGE message type),
but waitForEvent returned immediately (checked more message types).
Investigation shows that the difference was introduced in
https://github.com/debauchee/barrier/commit/dbfb04a6e
to fix a problem with bad behaviour of GetQueueStatus
Researching showed that a similar problem was fixed in Qt,
and the solution was
"pass different flags to GetQueueStatus depending on version of windows"
https://bugreports.qt.io/browse/QTBUG-29097
So this patch makes changes to a barrier non-GUI core similar to Qt fix.
QLength() may return 0 even if there are events pending because they
need to be read from the display socket in order to become visible. We
must use XPending() which will poll the socket if QLength() == 0.
Current code base don't use IEventQueue::isEmpty() method.
Remove it to simplify IEventQueue API and to remove confusion with
IEventQueueBuffer with same name.
The IEventQueueBuffer::isEmpty() and all its implementations kept unchanged
This changes the log level for several statuses (started server,
connected to server, server status:) from CLOG_NOTE or CLOG_INFO to
CLOG_PRINT so they will be printed to stdout regardless of the log
level. This allows the GUI to accurately report the status of the
`barriers` or `barrierc` processes in src/gui/src/MainWindow.cpp#L379-L399.
Errors where the socket address is in use will keep the server from connecting
with clients. This should be reported as an ERROR instead of a warning. Also,
if the retry time is shorter issues will be more obvious, since the error will
cause the log to scroll visibly.
Sometimes -lpthread is not enough. This has caused build failures on
Gentoo Linux in the past.
The logic also seems needlessly complex. We check for HAVE_PTHREAD on
UNIX platforms even though CMake errors out if pthreads are not
found. Similarly, we have a fallback for HAVE_PTHREAD_SIGNAL being
false even though we always set it to true.
The commit a841b28 changed the condition for removing job from processing.
New flag MultiplexerJobStatus::continue_servicing become used
instead of checking pointer for NULL.
However for cases when TCPSocket::newJob() returns nullptr
the behaviour changed: earlier the job was removed, but after change
it is called again, since MultiplexerJobStatus equal to {true, nullptr}
means "run this job again".
This leads to problem with eating CPU and RAM on linux
https://github.com/debauchee/barrier/issues/470
There is similar windows problem, but not sure it is related.
https://github.com/debauchee/barrier/issues/552
Since it looks that the goal of a841b28 was only clarifying
object ownership and not changing job deletion behaviour,
this commit tries to get original behaviour and fix the bugs above
by returning {false, nullptr} instead of {true, nullptr}
when TCPSocket::newJob() returns nullptr.
Member initializers are always run in the order defined in the header.
This commit reorders them in the implementation so that their order
reflects their execution order.