ui: fix VNC client throttling when forced update is requested

The VNC server must throttle data sent to the client to prevent the 'output'
buffer size growing without bound, if the client stops reading data off the
socket (either maliciously or due to stalled/slow network connection).

The current throttling is very crude because it simply checks whether the
output buffer offset is zero. This check is disabled if the client has requested
a forced update, because we want to send these as soon as possible.

As a result, the VNC client can cause QEMU to allocate arbitrary amounts of RAM.
They can first start something in the guest that triggers lots of framebuffer
updates eg play a youtube video. Then repeatedly send full framebuffer update
requests, but never read data back from the server. This can easily make QEMU's
VNC server send buffer consume 100MB of RAM per second, until the OOM killer
starts reaping processes (hopefully the rogue QEMU process, but it might pick
others...).

To address this we make the throttling more intelligent, so we can throttle
full updates. When we get a forced update request, we keep track of exactly how
much data we put on the output buffer. We will not process a subsequent forced
update request until this data has been fully sent on the wire. We always allow
one forced update request to be in flight, regardless of what data is queued
for incremental updates or audio data. The slight complication is that we do
not initially know how much data an update will send, as this is done in the
background by the VNC job thread. So we must track the fact that the job thread
has an update pending, and not process any further updates until this job is
has been completed & put data on the output buffer.

This unbounded memory growth affects all VNC server configurations supported by
QEMU, with no workaround possible. The mitigating factor is that it can only be
triggered by a client that has authenticated with the VNC server, and who is
able to trigger a large quantity of framebuffer updates or audio samples from
the guest OS. Mostly they'll just succeed in getting the OOM killer to kill
their own QEMU process, but its possible other processes can get taken out as
collateral damage.

This is a more general variant of the similar unbounded memory usage flaw in
the websockets server, that was previously assigned CVE-2017-15268, and fixed
in 2.11 by:

  commit a7b20a8efa28e5f22c26c06cd06c2f12bc863493
  Author: Daniel P. Berrange <berrange@redhat.com>
  Date:   Mon Oct 9 14:43:42 2017 +0100

    io: monitor encoutput buffer size from websocket GSource

This new general memory usage flaw has been assigned CVE-2017-15124, and is
partially fixed by this patch.

Signed-off-by: Daniel P. Berrange <berrange@redhat.com>
Reviewed-by: Darren Kenny <darren.kenny@oracle.com>
Reviewed-by: Marc-André Lureau <marcandre.lureau@redhat.com>
Message-id: 20171218191228.31018-11-berrange@redhat.com
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>

1 files changed, 7 insertions, 0 deletions

diff --git a/ui/vnc.h b/ui/vnc.h
index 8fe69595c6..3f4cd4d93d 100644
--- a/ui/vnc.h
+++ b/ui/vnc.h
@@ -271,6 +271,7 @@ struct VncState
 
     VncDisplay *vd;
     VncStateUpdate update; /* Most recent pending request from client */
+    VncStateUpdate job_update; /* Currently processed by job thread */
     int has_dirty;
     uint32_t features;
     int absolute;
@@ -298,6 +299,12 @@ struct VncState
 
     VncClientInfo *info;
 
+    /* Job thread bottom half has put data for a forced update
+     * into the output buffer. This offset points to the end of
+     * the update data in the output buffer. This lets us determine
+     * when a force update is fully sent to the client, allowing
+     * us to process further forced updates. */
+    size_t force_update_offset;
     /* We allow multiple incremental updates or audio capture
      * samples to be queued in output buffer, provided the
      * buffer size doesn't exceed this threshold. The value