Jakub Janků authored Jul 03, 2019 and Frediano Ziglio committed Jul 03, 2019

The current approach with OutputQueue in webdav has several problems:

* if the connection is slow, webdav keeps reading from phodav
and pushing messages to the internal channel xmit_queue.
This way, the queue can grow very quickly and the whole file
that is being transferred using webdav essentially gets loaded into memory.

* spice channel first flushes all messages in the xmit_queue and
then proceeds to reading. If webdav floods the xmit_queue with
a ton of messages, spice channel does not leave iterate_write until
the queue gets empty. This way, reading from the channel is blocked
till the whole file is transferred.

* OutputQueue uses g_output_stream_flush_async() on SpiceVmcOutputStream
that does not implement flush

To solve these issues, don't read from phodav until the last message
for a given client is written out to the socket.
(main channel currently uses the same approach when transferring files)

OutputQueue used an idle function to schedule the write and then
called mux_pushed_cb which started reading from phodav with priority
G_PRIORITY_DEFAULT.
Since this new approach does not utilize the idle scheduling,
lower the priority in client_start_read() to G_PRIORITY_DEFAULT_IDLE
to make sure other sources with lower priority get dispatched as well
(e.g. signals from coroutines, redrawing and resizing operations).

Also implement spice_webdav_channel_reset(). This is necessary because
spice_vmc_write_async() references the channel. If the channel is to be
disconnected, the write operations need to be cancelled so that
the references to the channel are released asap. Otherwise,
spice session would be stuck waiting for the channel to finalize.

Signed-off-by: Jakub Janků <jjanku@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>