SSL error occurs in non-main thread
This commit reworks how SSL errors are handled in QGIS. Previously,
ssl errors emitted by non-main thread QgsNetworkAccessManager instances
were handled by the main thread in QgisApp on next event loop. But
app (in the main thread) tried to pause the timeout timer for the
QNetworkReply containing the error, which doesn't work, because you
can't stop a timer in the main thread for a timer object created
in another thread. This meant that the reply could get deleted in
the background thread while the main thread was still using it
and waiting for users to respond to the SSL error. Now the timer
handling is done in the background thread's network access manager
instance, so in the thread matching the reply's timeout timer. We
then have to do some fancy thread locking using QWaitCondition,
because we need to "pause" that thread until the SSL error
is handled in the main thread -- if we don't pause the background
thread, Qt immediately resumes the QNetworkReply without ever
giving us the change to ignore ssl errors in the reply. Phew!
Additionally, the previous approach had a shortcoming in that
there was no way to notify background threads that ssl errors
had actually be handled. To do this we need a new signal which
can be fired after app has shown the ssl dialog and given users
a chance to respond. BUT... we can't safely do this -- if we
add a method to notify background threads when ssl errors have
been handled, then it CANNOT safely reside in app -- doing so
would break for QGIS server and QField etc, where theres no
app instance around to provide this notification. As a result
I've abstracted out the ssl error handling. By default there's
a simple error handler which just logs errors and returns
without ignoring them (i.e. default Qt behavior, an ssl error
cancels the request). App has a specific subclass of the ssl
error handler which presents the nice dialog and asks users to
choose what to do. Potentially server could decide to make
its own subclass too, which could e.g. ignore SSL warnings
if an environment variable is present (but at the moment
the behavior is effectively unchanged for server).
The end result is that SSL error handling should now be
totally thread safe, and we shouldn't hit any more deadlocks
and crashes when ssl errors occur in background threads.
I've also taken the opportunity to add a new signal which
is always emitted by the main thread QgsNetworkAccessManager
instance, which is emitted whenever ANY request on any
thread encounters an SSL error, and contains the requestId
so that the ssl error can be linked back to the originating
request. This is for debugging, and for use by the
network monitoring plugin to show ssl errors encountered
by requests.
