[PATCH 0/5] A couple more present fixes

[Keith Packard]

I've gotten a couple more bugs in Present resolved, and I'd love to
have some review of the current set of five tiny patches before I
merge them to master. Three of these are a repost, the last two are
new

Again, I'm soliciting any kind of comments as I'd like to get these
into the server before I release another RC on Friday.

-keith

 [PATCH 1/5] present: Round fake MSC computations intead of truncating

I don't know if anyone has noticed, but Present doesn't suffer the
DRI2 problem of running apps at infinite speed when DPMS fires, or
when the windows go off-screen. Instead, it has a 'fake' MSC timer
that keeps them going at a sensible speed.

For drivers with a real MSC interrupt available, the fake MSC runs at
1Hz. This means that when your screen blanks, all of your applications
slow down to one update per second, saving a tonne of power.

For drivers without a real MSC, the fake MSC runs at 60Hz. This means
that there is no driver support required for a usable Present
extension; you get updates at a reasonable rate on all cards, although
you won't avoid tearing unless you have hardware support.

In any case, this patch cleans up the computation of the fake MSC
value (it's computed from GetTimeInMicros, after all), to round rather
than truncate, which avoids having the MSC go backwards.

 [PATCH 2/5] present: Ignore event_id 0 from driver vblank notify

This is a trivial patch to work around potential driver bugs; if a
driver delivers a vblank event_id of zero, it "can't" be valid, and
may accidentally match sentinal values in various structures.

 [PATCH 3/5] present: Signal destroyed flip window with vblank->window

This changes some internal workings of the Present code to use a NULL
window value rather than a separate boolean for tracking when the flip
window has been destroyed (and hence the flip pixmap needs to be
'unflipped' asap). This eliminates some bugs where the code was
dereferencing through the flip window pointer after it had been freed.

 [PATCH 4/5] present: Block for wait_fence in present_execute

This adds a bit of missing functionality -- the server wasn't waiting
for the provided 'wait_fence' to be signaled before presenting the
pixmap.

 [PATCH 5/5] present: When unflipping, copy to flip window rather than

This avoids spamming old window contents on top of other windows on
the screen. I discovered this could occasionally happen when starting
a window manager with a full-screen GL application running along with
other apps. One of the intermediate stages while the window manager
was starting had the GL window full-screen and uncovered, causing a
page flip, and then immediately covering that with a decorated client
window; the window background paint would happen and then the whole
screen would get painted with the old GL window contents. This was
easily fixed by painting the old back buffer contents into the client
window itself, which respected the new clip list.