03:44imirkin: Lyude: friendly ping on the mst thing
08:11pabs3: my GPU seems to be in some sort of weird freeze state. GL apps freeze but X11 works fine. compositing WMs seem to freeze too. any idea how to debug this?
08:11pabs3: when I press Ctrl+C in gdb, I get this: http://paste.debian.net/1025193/
08:13pabs3: aha, I also get this in dmesg: http://paste.debian.net/1025194/
08:14pabs3: this is with Linux 4.16.5-1 from Debian buster
08:15pabs3: my GPU: 01:00.0 VGA compatible controller: NVIDIA Corporation GK107 [GeForce GT 740] (rev a1)
09:38karolherbst: skeggsb, imirkin: I am currently thinking if we should disable runpm support on all laptops doing this _PR thing, because it is broken on 2 laptops I have access to out of 2 having _PR
09:39karolherbst: I am not talking about disabling it inside nouveau, but on the whole kernel level, because I am sure it is broken for AMD GPUs as well
09:57karolherbst: I will try to figure something out on this new laptop I got my hands on, which has a gm204 GPU
10:05airlied: karolherbst: sounds like the sort of problem RH would want fixed
10:07karolherbst: the situation is.... weird
10:11karolherbst: anyway, I tried to look into it and well
10:11karolherbst: it's just a bug inside the linux kernel basically
10:11karolherbst: so basically what happens is, that the GPU suspends just fine
10:11karolherbst: also resuming just works, but we don't get any connection through the PCIe bus
10:12karolherbst: so what works is: remove PCIe device, call ACPI methods, rescan bus -> load nouveau
10:26karolherbst: ahhh github change their website layout :(
11:34imirkin: pabs3: update to a newer kernel. swiotlb error messages got fixed at some (recent) point ... finally.
11:36pabs3: imirkin: any idea which version? 4.16.9 or 4.17-rc5?
11:37imirkin: yes, either one of those
11:39karolherbst: imirkin: famous last words
12:19karolherbst: hum, I just spammed some glxspheres and this happened: https://gist.githubusercontent.com/karolherbst/3abe9eb280196529c38bdbc437d25607/raw/23795b29750b0bf0179ffc6062cf99b97d916023/gistfile1.txt
12:20karolherbst: I guess this is the famous multi context issue we hit?
12:36imirkin: karolherbst: glxspheres shouldn't be threaded afaik
12:36imirkin: [ 513.452930] nouveau 0000:02:00.0: gr: GPC0/PROP trap: 00000400 [RT_LINEAR_MISMATCH] x = 368, y = 66, format = 2a, storage type = 0
12:36imirkin: that's the real issue...
12:36imirkin: some kind of buffer mismanagement
12:38imirkin: (or context switch issue)
12:38karolherbst: or that
12:38karolherbst: I was running it on a gk106 which should have those issues where blob firmware helps
12:39karolherbst: but this something I would like to look into next week
12:39karolherbst: I was just surprised that I triggered that after a few seconds
12:39imirkin: yeah, for those GK106's, the issues come about quickly.
12:39karolherbst: I see
12:39karolherbst: so I guess something in the firmware is fishy then and maybe I can track it down with that card
13:13imirkin_: or something with how gr is initialized
13:13imirkin_: which their ctxsw handles but ours doesn't
13:13imirkin_: check ben's recent stuff
13:14imirkin_: he has some gr init updates
13:14imirkin_: not sure if they'd matter
13:34karolherbst: I ran it on his master branch
13:35karolherbst: I also gave him traces of this card, but he said he didn't catch anything suspicious
19:51diogenes_: Hello guys, so i've got nvidia optimus (intel + nvidia), i use nouveau and if i understand it right, in order to run an application with nvidia card, i have to use: DRI_PRIME=1 application, but is there any monitoring tool something similar to nvidia-settings to see the gpu percentage utilization and stuff?
19:58imirkin_: there is not.
20:01HdkR: tsk tsk :P
20:01imirkin_: send patches.
20:01imirkin_: i haven't a clue how to even measure gpu utilization
20:02diogenes_: imirkin_, so how can one be sure if the gpu is being utilized?
20:02imirkin_: DRI_PRIME=1 glxinfo
20:02imirkin_: if that says nouveau, then you're good.
20:02imirkin_: you can also tell based on whether nouveau_dri.so is loaded (check /proc/<pid>/maps)
20:03imirkin_: i suppose we could expose various counters, like "job submitted" and so on
20:03imirkin_: (like perf events or something)
20:03glennk: doesn't nouveau expose various counters for GALLIUM_HUD?
20:03diogenes_: imirkin_, ok thanks, one more question, are there any tricks to improve the performance with nouveau?
20:04imirkin_: glennk: sure, but this is for a side application
20:04imirkin_: diogenes_: what GPU?
20:08diogenes_: imirkin_, NVIDIA Corporation GK107M [GeForce GT 650M]
20:09orbea: before xorg 1.20.0 DRI3 + modesetting had a small, but noticable perf boost, now it doesn't work...
20:10orbea: reclocking too ofc
20:13imirkin_: diogenes_: assuming you're on kernel 4.10 or later, you can reclock your gpu
20:13imirkin_: it boots into the lowest perf level, so you should be able to get a lot more speed out of it
20:13imirkin_: cat /sys/kernel/debug/dri/0/pstate (or /1/pstate)
20:14imirkin_: that should give you a list of the available perf levels
20:14diogenes_: imirkin_, i've just done the furmark test and i got with intel modesetting - 274 points and with nouvea - 184
20:14diogenes_: kernel 4.4
20:14imirkin_: ah yeah, with kernel 4.4 your options are limited
20:14imirkin_: there was early reclocking then, i think ... you may be able to go up to a middle clock speed
20:15imirkin_: 4.4 came out like 2y ago though, so you're missing out on some fixes :)
20:15orbea: I'd think updating the graphics stack/kernel might be worth it in general for nouveau :)
20:15diogenes_: imirkin_, ok, soon my distro will be upgraded to 1.16 i guess
20:15imirkin_: hopefully 4.16 :)
20:16diogenes_: 4.16 :)
20:16imirkin_: what intel is in there? haswell?
20:17imirkin_: i'm not surprised that on lowest perf level it wouldn't have outperformed
20:17diogenes_: imirkin_, i7 ivi bridge
20:17imirkin_: ivybridge was pretty slow
20:18diogenes_: maybe but in furmark it gets mre points than my nvidia
20:18imirkin_: with the lowest perf level, not super-surprising
20:18imirkin_: cat that pstate file
20:18imirkin_: should tell you where you are now, and what the card max is
20:18imirkin_: (pastebin the results)
20:18diogenes_: imirkin_, sorry, i missed something, what should i cat?
20:18imirkin_: oh, maybe in 4.4 it was actually still in sysfs, not debugfs yet
20:19imirkin_: so long ago
20:19imirkin_: 4.4 released Jan 10, 2016
20:20diogenes_: imirkin_, here is the output of cat: http://termbin.com/i44pk
20:20imirkin_: ok right. so it's off now, hence the zero's
20:20imirkin_: my guess is that you should be able to reclock to 0a in kernel 4.4
20:21imirkin_: if you feel like trying it, start something that uses the gpu
20:21imirkin_: and then as root
20:21imirkin_: echo 0a > /sys/kernel/debug/dri/0/pstate
20:21imirkin_: note that there's a 50% chance that your comp will hang.
20:21imirkin_: so save your work :)
20:23imirkin_: lol. https://github.com/skeggsb/nouveau/issues/2
20:23diogenes_: imirkin_, i ran: echo 0a > /sys/kernel/debug/dri/0/pstate but nothing's going on, only the fans are louder now
20:23imirkin_: diogenes_: while an app was using the gpu?
20:23imirkin_: (the nvidia gpu)
20:23diogenes_: imirkin_, nothing was using the gpu
20:24imirkin_: ok, so now you have a hung cpu probably
20:24imirkin_: like i said above, something must be running on the gpu
20:24imirkin_: so you should reboot.
20:24diogenes_: ok i reboot and come back
20:25imirkin_: skeggsb: the fps on intel is much more better!
20:26diogenes_: imirkin, so what should i do now? what are the steps?
20:27imirkin_: diogenes_: step 1: make sure something is running on nvidia gpu
20:27imirkin_: step 2: echo 0a > .../pstate
20:27imirkin_: step 3: enjoy higher perf
20:27diogenes_: ok let me see
20:27imirkin_: step 2 must happen WHILE osmething's urnning on the gpu
20:27imirkin_: also i think once it stops, the settings won't persist
20:27diogenes_: so i have to run: DRI_PRIME=1 application
20:27imirkin_: (the gpu gets auto-suspended when not used)
20:27diogenes_: then run the echo
20:27imirkin_: while the app is running
20:28diogenes_: ok let's see :)
20:28imirkin_: note that 4.4 is a long time ago, and we had a lot of trouble with reclocking in those days
20:28imirkin_: 0f definitely won't work out
20:28imirkin_: 0a might work
20:28imirkin_: with a 4.10+ kernel 0f should work
20:29imirkin_: (v4.10 was released Feb 19, 2017)
20:31diogenes_: imirkin_, wow, i actually got twice as much as intel gave me, so it's definitely cool, how can i make it persistent?
20:32imirkin_: you can't.
20:32imirkin_: er wait... no you can
20:32imirkin_: add nouveau.config=NvClkMode=10 to your kernel config
20:32diogenes_: how how how :)
20:32diogenes_: to grub?
20:32imirkin_: (or equivalent effect via modprobe config)
20:32imirkin_: er, add ... to your kernel cmdline, of course. not config. my bad.
20:32diogenes_: let's see :) thanks so much for such a cool trick
20:35diogenes_: imirkin_, ok added it and before i reboot, one more little question, when i boot, i notice a few messages like this:
20:35diogenes_: nouveau 0000:01:00.0: iccsense: Unknown sensor type 30, power reading disabled
20:35imirkin_: you should get another 4x improvement with 0f, i suspect.
20:36imirkin_: that's fine
20:36imirkin_: someone might be interested in your vbios though
20:37diogenes_: so what do you say, instead of NvClkMode=10 i should add NvClkMode=0f?
20:37imirkin_: =15. but that won't work with kernel 4.4
20:37diogenes_: ok, let me reboot and see what's going on
20:43diogenes_: imirkin, yay! perfect, it works like a charm, thank you very much and i think i could add this trick to archwiki because it's not mentioned anywhere and that's not right, everyone should know this, nouveau is cool!
20:57imirkin_: diogenes_: not according to this guy: https://github.com/skeggsb/nouveau/issues/2
21:00imirkin_: apparently nouveau fps must be much more!
21:01diogenes_: imirkin_, well with the trick you provided, now it's 2x more than intel
21:02diogenes_: previously it was much lower than intel
21:45diogenes_: imirkin, i've just tried with =15 and it's amazing, i got a result very close to the nvidia proprietary driver, so it's trully amazing!
21:45diogenes_: so it turns out, it works on 4.4
21:46HdkR: "Nvidia devs hate him using this one simple trick. Click to find out more."
21:48annadane: what are people talking about?
21:48diogenes_: about improving nouveau performance
21:48HdkR: putting the GPU in a higher power mode to get better perf
21:49annadane: i kind of just caved and started using the proprietary driver