01:00 karolherbst: Lyude: sooo, using reset_function does seem to improve the situation as in lspci is happy... but using glxinfo works only like once
01:00 karolherbst: could be some secboot messup on top of it though
01:01 karolherbst: but my machine hard locks
01:18 Lyude: Huh
01:19 Lyude: Didn't you mention something about the link rate changing when the secboot fail happened?
01:22 Lyude: (@ karolherbst)
01:24 karolherbst: Lyude: devinit changes the link rate
01:26 Lyude: where are you doing the reset by the way?
01:26 karolherbst: nouveau_pmops_runtime_suspend
01:30 Lyude: and we made sure we can't call the reset on resume?
01:31 karolherbst: maybe we have to do both?
01:31 Lyude: Worth a shot
01:35 Lyude: Probably reset the device as late as possible when suspending, then reset it as early as possible when resuming
02:19 karolherbst: Lyude: pci_disable_device is already called by the PCI subsystem, no?
02:20 karolherbst: mhh, actually. I don't think so
02:24 karolherbst: Lyude: sooo, I think it's an improvement, but... when secboot was performed then it still fails
02:25 karolherbst: can't debug as even the SSH session is dead
02:26 karolherbst: but without triggering secboot, I can runtime suspend/resume the GPU multiple times
02:27 karolherbst: essentially the same behaviour as with my pci link hack
02:33 Lyude: Netconsole maybe?
02:51 karolherbst: Lyude: I think if ssh dies, the entire kernel is essentially dead
02:52 karolherbst: but yeah... something weird is going on, why would the machine just die if the GPU can't be resumed
02:55 karolherbst: Lyude: but seriuosly, I think we have to wait until nvidia actually implements it themselves, then we can reverse engineer it, fix nouveau, and then after we've got the docs, we can check if we actually got it right :p
02:55 Lyude: karolherbst: it is, but you can usually get more information with netconsole
02:56 Lyude: It's literally printk wired to your Ethernet jack
02:56 karolherbst: well, the last line I've got was "refused to change power state, currently in D3"
02:56 karolherbst: "dmesg -w" over ssh is still quite powerfull ;)
03:13 Lyude: it is!
03:13 Lyude: I use it most of the time
03:14 Lyude: But when I'm not getting output, I have a script I use that sets up netconsole for me and I've gotten more output that way quite a few times
03:15 Lyude: not having to go through user space at all makes printk a lot more likely to work
04:41 karolherbst: imirkin: with MESA_DEBUG=flush I don't get the ctxsw timeout :( *sigh*
04:42 imirkin: heh
04:42 imirkin: yay timing!
04:43 karolherbst: slowly I think this is just some super stupid issue in the end
17:31 karolherbst: imirkin: any other ideas besides MESA_DEBUG=flush? :/
17:31 imirkin_: i know how to make it flush *harder*
17:31 imirkin_: but not sure about less hard :)
17:34 karolherbst: mhhh
17:36 karolherbst: but why does kicking the pb more often would change something here. Can only think of some fencing stuff or something
17:37 imirkin_: or happened to trip it a bit
17:37 imirkin_: er, move things around
21:53 Lyude: skeggsb: poke: I notice with the nv50_outp_acquire() methods we pass both the or index and something called "link", what is that?
22:32 rawf: I've seen claims that the GTX 780 can run 3D without nvidia's firmware. This contradicts the feature matrix on the nouveau website. Does anyone know if this card an exception in the NVE0 family or is this a false claim?
22:43 imirkin_: rawf: what bit of the feature matrix are you referring to?
22:50 rawf: woops, nevermind :|
23:12 imirkin_: glad i could help :)
23:15 Lyude: skeggsb: poke, if you've got a moment I've got some more link training related questions for nouveau
23:15 imirkin_: rawf: note we still lag considerably behind nvidia in terms of perf, even with the reclocking situation sorted out (as it is on kepler)
23:37 imirkin_: should probably update that feature matrix...