00:03 gnarface: brief googling suggests it might be a problem with power management
00:06 Lyude: pabs3: should file a bug on freedesktop
00:07 Lyude: also hooray, my gpu collection now goes further back then fermi: all the way back to curie
00:20 Lyude: looks like curie has some memory access bugs though
00:22 Katana_Steel: I seem to recall OS reboots using my ASUS 6600GT back in the day using the other OS
00:28 Katana_Steel: I think I still have the 6600GT card somewhere :)
00:58 pabs3: Lyude: is that going to be useful considering it was completely random and I have no way to reproduce it?
00:59 imirkin: Lyude: hm?
01:03 Lyude: pabs3: maybe, maybe not. it's good to have it on record though in case someone else hits the same bug
01:04 Lyude: imirkin: the Curie card I've got now causes a good number of unhandled page faults with the iommu on init and general usage
01:04 Lyude: Seems to still work though
01:05 Lyude: But we're definitely touching unmapped memory
01:06 skeggsb: Lyude: any particular address? or random?
01:08 Lyude: 0x0, 0x200, 0x204 are the ones I remember seeing. A couple of others as well, but I won't be able to check the machine again until tommorrow
01:09 Lyude: Didn't seem to be too random
02:57 imirkin: Lyude: which one's curie? nv4x?
02:58 imirkin: i have a nv34 plugged in, seems to be fine
04:10 imirkin: skeggsb: did you have any ideas about the WARN's in nvif_vmm_put ?
04:11 imirkin: Lyude: oh wait, i remember nv4x hitting a bunch of low addresses if you have iommu enabled. iirc that's "expected"
04:11 skeggsb: imirkin: if you can find me a solid way to reproduce them, i might be able to get an idea :P
04:11 imirkin: skeggsb: sure thing. sit down in front of my computer, and use it for a day. check dmesg, and there they are :)
04:11 skeggsb: haha ;)
04:12 imirkin: the GPU i'm currently using is a GTX 960 (GM206)
04:12 imirkin: same no-graphics-heavy-anything environment i always use, except chrome
04:13 imirkin: i did try to run the khronos webgl tests in chrome
04:13 imirkin: which eventually crashed the webgl context
04:13 skeggsb: did nvidia say those were expected on nv4x?
04:13 imirkin: and spat some interesting DATA_ERROR's
04:13 imirkin: skeggsb: i don't think so. but iirc they also happened with their driver.
04:13 imirkin: this is very vague recollection though
04:14 imirkin: https://bugs.freedesktop.org/show_bug.cgi?id=73233
04:14 skeggsb: ie. are they read faults used to "flush" sysmem accesses
04:15 skeggsb: see nvkm/subdev/fb/nv50.c::nv50_fb_init()
04:15 skeggsb: g80 and up you point at a real page, perhaps nv4x just accesses some random addresses
04:15 skeggsb: *or* it's configurable and both us and RM don't
04:16 imirkin: well, based on that bug report, nvidia driver actually worked ok
04:16 imirkin: so my recollection is wrong
04:16 imirkin: or i was thinking of a different report
04:16 imirkin: which does not match "nouveau iommu error" search :)
04:21 imirkin: skeggsb: so ... you don't see that nvif_vmm_put warning locally at all? hrm
04:21 imirkin: both karol and i have seen it
04:22 skeggsb: i basically never reboot... let me check dmesg
04:22 skeggsb: not on my gk107 laptop anyway
04:23 imirkin: well, depending how often 'never' is... this didn't use to happen in like 4.12 :)
04:23 imirkin: but it does happen in 4.18.5
04:32 skeggsb: never is basically, when i'm forced to update the kernel for drm-next
04:32 skeggsb: skeggsb@nisroch ~ $ uname -r
04:32 skeggsb: 4.19.0-rc2+
04:32 imirkin: hehe
04:34 skeggsb: i've not seen it on volta either while running piglit 150,000,000 times :P
04:35 skeggsb: and i generally have a window scrolling dmesg open when that's running
20:52 Lyude: btw danvet_ yeah i'm aware of the vcpi stuff with nouveau
20:52 Lyude: there's other things we do wrong in mst as well, iirc we don't default to the highest available link bandwidth
21:07 danvet_: Lyude, amdgpu seems to have the same yolo vcpi allocation :-)
21:07 Lyude: danvet_: but it's even worse! none of it even happens in the atomic checks, lol
21:07 Lyude: (that's the stuff I was referring to at XDC btw)
21:32 danvet_: Lyude, yeah that's what I meant with yolo vcpi allocation, it never reports an error to userspace
21:32 danvet_: not very atomic
21:35 Lyude: i bet you there are many many other spots like that in DC/DM we haven't noticed yet, tbh
21:35 danvet_: yeah ...