[08:43] <duflu> Ha. Not a bug.
[08:43]  * duflu throws out broken mouse
[08:43]  * alan_g discovers that "sudo unity-system-compositor --debug-without-dm" crashes horribly
[08:51] <anpok> duflu: get a cat!
[08:51] <anpok> I can make the seg fault on the krillin ci runs go away
[08:51] <duflu> Don't need one. Have the internet full of cats.
[08:52] <anpok> it seems that.. if we dont load the mesa/kms platform unloading the android platform does not crash the driver
[09:09] <alan_g> anpok: that is likely the "horrible hack"
[09:10] <anpok> so..
[09:10] <anpok> there are various ways
[09:10] <alan_g> The mesa driver code reloads itself with RTLD_GLOBAL
[09:10] <anpok> oh
[09:10] <anpok> i thought some tls related problems..
[09:11] <anpok> another option might be to never unload the android platform if quirk->cannot_survive_unloading()..
[09:13] <anpok> or change the ci script to handle krillins runners differently.. disabling the probe .. removing the mesa modules .. all of which harms the effectifness of ci
[09:14] <alan_g> It is just on krillin?
[09:14] <anpok> yes
[09:15] <alan_g> A thought: what if you probe the android drivers before the mesa ones?
[09:15] <anpok> hm that is happening
[09:16] <alan_g> Rats!
[09:17] <anpok> brb
[09:19] <alan_g> anpok: for differential diagnosis, could you try the effect of commenting out the reload code - ensure_loaded_with_rtld_global() - in the mesa drivers?
[10:04] <anpok> alan_g: bingo
[10:05] <anpok> when do we need that?
[10:05] <anpok> could we delay that until we know this will be the platform to use?
[10:06] <anpok> arg thats in the init code of the library
[10:06] <alan_g> anpok: somewhere in the mesa code there's an attempt to find some symbol. (I think alf knows the details.)
[10:06] <anpok> oh or we might be able to split the mesa driver into pieces..
[10:06] <anpok> since probing only needs udev..
[10:07] <alan_g> anpok: RAOF promised to have a go at fixing mesa (eventually) as it also pulls in the X stack and other rubbish
[10:11] <alan_g> So delaying the hack is a possibility - depending on when mesa needs to find ... mir_server_mesa_egl_native_display_is_valid()
[10:13]  * alan_g wonders if there's a way to just publish that one symbol...
[11:51] <anpok> alan_g: we only need that in nested, right?
[11:52] <alan_g> anpok: no for any mesa graphics (AIUI)
[13:26] <anpok> vogons: https://code.launchpad.net/~andreas-pokorny/mir/fix-1506137/+merge/274708
[13:27] <anpok> alan_g: it seems to work for the guest platform only - which makes sense since the drm-egl is builtin mesa and needs no mir interaction
[13:32] <alan_g> anpok: if that's true then great. (Not as good as getting rid of the hack, but at least it reduces the cases where it messes us up).
[13:34] <anpok> i still have vt switching issues.. so I tried it via two times cascaded launch-client... and it seemed to work
[13:34] <anpok> oh
[13:34] <anpok> thats not the right use case..
[13:35] <anpok> I should have tried it with unity8 instead?
[13:37] <anpok> but still it should be true..
[15:13] <alf> robert_ancell: Hi!