Mir mesa-kms driver crashes ("Failed to schedule page flip") if you turn the monitor off

[Daniel van Vugt]

Mir crashes if you turn the monitor off.

Test case:
   1. Start a Mir server
   2. Turn the monitor off

Expected: Mir keeps running.
Observed: Mir crashes:

ERROR: /home/dan/bzr/mir/trunk/src/common/fatal/fatal.cpp(55): Throw in function void mir::fatal_error_except(const char*, ...)
Dynamic exception type: boost::exception_detail::clone_impl<boost::exception_detail::error_info_injector<std::runtime_error> >
std::exception::what: Failed to schedule page flip

[Daniel van Vugt (vanvugt)]

I always suspected "Failed to schedule page flip" was a soft error. It should not be treated as fatal.

[Alan Griffiths (alan-griffiths)]

I don't know if it is connected but (since late yesterday - I didn't try it earlier) I've been seeing that error on Wily when starting a server: No need to turn off a monitor.

It also happens with 0.14 - so possibly not a Mir change?

[Daniel van Vugt (vanvugt)]

Might be the new i915 kernel behaviour, but regardless I've always feared this would happen one day. We shouldn't treat it as fatal. Just ignore and carry on.

[Alan Griffiths (alan-griffiths)]

A bit more information:

I typically start a server with "--test-client <whatever>" and that seems to be what provokes the problem I see. If I start the server, then VT switch to start a client everything works. (Even if I switch a monitor off or on - so I can't confirm the above test case.)

This is with both trunk and 0.14 - so I agree in suspecting a change to kernel behaviour. While "carry on" seems plausible I think we ought to report that a problem occurred and not just ignore it.

No problems on vivid with either scenario or codebase.

[Alan Griffiths (alan-griffiths)]

OK filed my scenario as lp:1489806 - probably the same root cause, but keeping separate until we can confirm that.

[Alan Griffiths (alan-griffiths)]

 looking at our schedule_page_flip() code...

    for (auto& output : outputs)
    {
        if (output->schedule_page_flip(bufobj->get_drm_fb_id()))
            page_flips_pending = true;
    }

...if there are no outputs we don't set page_flips_pending - and that leads immediately to a fatal error.

That isn't the problem in my similar scenario, but might be the problem here (but I can't reproduce).

[Daniel van Vugt (vanvugt)]

That's a good observation but apparently benign. The class in question never gets instantiated if there are zero outputs.

[Daniel van Vugt (vanvugt)]

On second thoughts, this is quite messy. We've got more code that needs fixing than I expected. Another day...

[Daniel van Vugt (vanvugt)]

I'm struggling to reproduce this now, but it has been over 14 months. Regardless, the offending code has been removed in lp:~vanvugt/mir/fix-1584894.

[Mir CI Bot (mir-ci-bot)]

Fix committed into lp:mir at revision None, scheduled for release in mir, milestone 0.25.0

[Launchpad Janitor (janitor)]

[Expired for Mir because there has been no activity for 60 days.]