[PATCH] Re-enable the RECORD extension. (#20500)

[Peter Hutterer]

On Wed, Jan 13, 2010 at 02:29:46AM +1100, Daniel Stone wrote:
> Hi,
> 
> On Mon, Jan 11, 2010 at 04:13:25PM +1000, Peter Hutterer wrote:
> > On Sat, Jan 09, 2010 at 03:45:08AM +1100, Daniel Stone wrote:
> > > So it looks like we will record one Xi and one core event for every
> > > event that ends up getting sent through the MD, regardless of which
> > > deliveries were made.
> > > 
> > > Previously we'd only record an input event if any deliveries of that
> > > type occurred, right? It looks like something like the attached
> > > (note: wholly untested, had to peg my CPU at 12% to prevent thermal
> > > shutdowns ... MacBook Air + Australian summer is a poor combination)
> > > would match the old semantics.
> > 
> > Not quite. RECORD differs between "device events" and "delivered events".
> > The former include all events regardless of whether they are delivered so
> > the process here is correct. The naming in the code is somewhat unfortunate,
> > it's "device events" and "events" when they are really "all events" and
> > "delivered events".
> > 
> > Record requires that _any_ event is delivered, hence the hook in
> > WriteEventsToClient which is called for all events including randr, XI, etc.
> > 
> > It looks like Chris' original patch was correct anyway once the EventToCore
> > patch is applied. I'm still waiting for more test feedback.
> 
> Yeah, just checked the spec and you're right.  Aside from the two
> ErrorFs left in there, it seems like it behaves incorrectly for events
> queued while a device is grabbed.  From my reading of the spec, it looks
> like a device event should be generated from EnqueueEvent, but not from
> ProcessOtherEvent when the device is thawed; delivered events should
> still be sent though.

I don't think the recording should be stopped. Reading from section 2.1.4,
it says "the recording of selected device events is not affected by server
grabs" (where in this context "server grab" means pointer/keyboard grabs,
not GrabServer grabs)

The ambiguous bit is "if a recording client selects both a device event and
delivered events that result from that device event, the delivered event are
recorded after the device events. In the absence of grabs, the delivered
events for a device event precede later device events."

I read the second sentence as the order of events for ungrabbed devices will
be "device delivered device delivered device delivered ...".
For a grab, the order may be "device device device delivered" or something
similar. The "later" is the key here, the delivered event for an event
precedes a device event for an event occuring _after_ the first event.

Checking with the server 1.3 sources that predate the input rework we
recorded device events from CoreProcessKeyboardEvent and
CoreProcessPointerEvent as well as ProcessOtherEvents. The first two are
gone now, so recording in POE seems correct - at least to maintain the same
behaviour.

Cheers,
  Peter