Bug 94250

Created attachment 121894 [details]
stack trace, without --enable-debug

Created attachment 121895 [details]
stack trace, with --enable-debug

Created attachment 121896 [details]
Xorg.0.log.debug, with --enable-debug

Hmm. First thought is that the kernel returned an event for an operation that failed, can you please quickly try with

diff --git a/src/sna/sna_dri2.c b/src/sna/sna_dri2.c
index fcfbd9d..7db6f61 100644
--- a/src/sna/sna_dri2.c
+++ b/src/sna/sna_dri2.c
@@ -1914,6 +1914,7 @@ sna_dri2_flip(struct sna_dri2_event *info)
                return false;
        }
 
+       assert(!info->queued);
        if (!sna_page_flip(info->sna, bo, sna_dri2_flip_handler,
                           info->type == FLIP_ASYNC ? NULL : info))
                return false;

and --enable-debug.

Ok, found one scenario that could explain the traces after a failed flip when undocking, so please test with:

commit 64b1b1f10da59f15a91141c9f76d7d09517f8ea8
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date:   Tue Feb 23 09:32:57 2016 +0000

    sna/dri2: Ensure the flipqueue is drained on pageflip failure
    
    If we fail to queue a flip for a CRTC, we attempt to restore the
    original mode. However, if the failure is on a second CRTC, the queued
    flip on the first will still complete causing us to process the event
    twice.
    
    References: https://bugs.freedesktop.org/show_bug.cgi?id=94250
    Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

When using the first patch (comment 4), the code did not exit at the new assertion, but at the one from before (from comment 3).

I think I saw it one time exiting with (but failed to save the log and core dump):
(EE) sna_dri2_schedule_flip:3078 assertion 'info->front == front' failed

And the new patch (comment 5) unfortunately did not fix the problem. After a couple of tries the code seems to exit at one of two assertions. I will add the Xorg logs and corresponding stack traces. One of the assertions is "'info->front == front' failed" from above, so I got the log and stack trace for this one.

Created attachment 121918 [details]
Xorg.0.log.1

Xorg log for assertion

(EE) sna_dri2_event_free:1663 assertion '!info->signal' failed

Created attachment 121919 [details]
stack_trace1.txt

stack trace for assertion

(EE) sna_dri2_event_free:1663 assertion '!info->signal' failed

Created attachment 121920 [details]
Xorg.0.log.2

Xorg log for assertion

(EE) sna_dri2_schedule_flip:3078 assertion 'info->front == front' failed

Created attachment 121921 [details]
stack_trace2.txt

stack trace for assertion

(EE) sna_dri2_schedule_flip:3078 assertion 'info->front == front' failed

(In reply to cs_gon from comment #7)
> Created attachment 121918 [details]
> Xorg.0.log.1
> 
> Xorg log for assertion
> 
> (EE) sna_dri2_event_free:1663 assertion '!info->signal' failed

That is not a huge issue, just a bit of debug code needs cancelling along the error path.

(In reply to cs_gon from comment #10)
> Created attachment 121921 [details]
> stack_trace2.txt
> 
> stack trace for assertion
> 
> (EE) sna_dri2_schedule_flip:3078 assertion 'info->front == front' failed

This is actually more of an issue. This should reduce the assert

commit 3593a2d18928f74ee470f824dc34b8b5b148ce2d
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date:   Tue Feb 23 14:36:10 2016 +0000

    sna/dri2: Reset front pointer on frame event across a modeset
    
    If the root window's pixmap is changed (e.g. to resizing the
    framebuffer) then an outstanding flip becomes invalid. The invalid flip
    is marked as having a stale front, and that triggers an assertion if the
    user then tries to schedule flip before the pending flip event is
    processed.
    
    References: https://bugs.freedesktop.org/show_bug.cgi?id=9425
    Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

but it is still a little bit of a puzzle how we did not drain the event queue first. That's probably explained by

commit 4cd43faa646e368624079b73b216f6546ede5c16
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date:   Tue Feb 23 14:40:39 2016 +0000

    sna: Flush the DRM event queue after a modeset
    
    Changing the mode will cause the DRM events(pageflips, vblanks) to
    be completed and the queue flushed. After applying the CRTC, drain the
    event queue.
    
    Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

The last two patches may have fixed it, or at least made things a lot better. 

I found out before, that disconnecting the external monitor from the docking station, instead of undocking the notebook, did cause this problem more frequently (maybe 75% of the time), and I was now unable to reproduce the issue when disconnecting the monitor.

Though during a lot of docking and undocking, I got this assertion one time:
(EE) sna_dri2_schedule_flip:3080 assertion 'info->front == NULL' failed

I will attach the Xorg log and stack trace again.

Created attachment 121925 [details]
Xorg.0.log.3

Xorg log for assertion:

(EE) sna_dri2_schedule_flip:3080 assertion 'info->front == NULL' failed

Created attachment 121926 [details]
stack_trace3.txt

stack trace for assertion

(EE) sna_dri2_schedule_flip:3080 assertion 'info->front == NULL' failed

After a bit more thought, the assertion was incorrect and we just expect the stale pointer and not NULL.

Sadly, I squashed the change with

commit d1672806a5222f00dcc2eb24ccddd03f727f71bc
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date:   Wed Feb 24 10:33:22 2016 +0000

    sna/dri2: Add active-scanout tracking to single CRTC flips