(In reply to comment #0)
> This is a regression from 3.12 to 3.13 (still present on nouveau master).
should have been 3.11 to 3.12.

A git bisect would be useful. You should be able to just run the bisect under drivers/gpu/drm/nouveau, i.e.

git bisect start v3.12 v3.11 -- drivers/gpu/drm/nouveau

(In reply to comment #2)
> A git bisect would be useful. You should be able to just run the bisect
> under drivers/gpu/drm/nouveau, i.e.
> 
> git bisect start v3.12 v3.11 -- drivers/gpu/drm/nouveau

$ git bisect log
# bad: [5e01dc7b26d9f24f39abace5da98ccbd6a5ceb52] Linux 3.12
# good: [6e4664525b1db28f8c4e1130957f70a94c19213e] Linux 3.11
git bisect start 'v3.12' 'v3.11' 'drivers/gpu/drm/nouveau/'
# bad: [bd9c5a2016307164c419c5e24a46921c10e620a0] drm/nouveau: require contiguous bo for framebuffer
git bisect bad bd9c5a2016307164c419c5e24a46921c10e620a0
# bad: [0d69704ae348c03bc216b01e32a0e9a2372be419] gpu/vga_switcheroo: add driver control power feature. (v3)
git bisect bad 0d69704ae348c03bc216b01e32a0e9a2372be419
# bad: [baa7094355a10b432bbccacb925da4bdac861c8d] drm: const'ify ioctls table (v2)
git bisect bad baa7094355a10b432bbccacb925da4bdac861c8d
# good: [72525b3f333de54fa0c42ef87f27861e41478f1e] drm/ttm: convert to unified vma offset manager
git bisect good 72525b3f333de54fa0c42ef87f27861e41478f1e
# bad: [43387b37fa2d0f368142b8fa8c9440da92e5381b] drm/gem: create drm_gem_dumb_destroy
git bisect bad 43387b37fa2d0f368142b8fa8c9440da92e5381b
# first bad commit: [43387b37fa2d0f368142b8fa8c9440da92e5381b] drm/gem: create drm_gem_dumb_destroy


note that the only good commit was a bit weird, plymouth won't run, gnome-shell segfaults immediatelly, and DRI_PRIME=13d apps showed no output (both glxgears and oa). However, both were running (displayed fps, ...) and the hang was always casued by running oa (never glxgears) so I figured the black screen does not matter.

Hmm... that commit is outside of nouveau. I'm not 100% sure that the bisect results are accurate in that case. (But all the work you did is still useful, future bisects can use the same good/bad information.)

Can you verify that 43387b37fa2d0^ is good? (i.e. commit 86e81f0e62)

If that is not the case, you can restart the bisect with

git bisect v3.12 v3.11 72525b3f333de54 -- drivers/gpu/drm

which will include everything in drm.

If oa does not hang, I also see a lot of:

[  683.090527] dmar: DRHD: handling fault status reg 3
[  683.090531] dmar: DMAR:[DMA Write] Request device [01:00.0] fault addr fe000000 
DMAR:[fault reason 05] PTE Write access is not set

There's 1400+ entries in generally ascending order (sometimes it goes back or stays on the same address). The last one is:

[  683.093190] dmar: DRHD: handling fault status reg 3
[  683.093192] dmar: DMAR:[DMA Write] Request device [01:00.0] fault addr fe579000 
DMAR:[fault reason 05] PTE Write access is not set

I'm not sure if it's a direct correlation (hang xor dma errors)

As for the DMAR errors, see if disabling the IOMMU (aka VT-d for intel) helps. (It'll definitely get rid of the messages, but might not get rid of the underlying issues.)

nvm, I just reproduced it on 3.11.6. It was a bit harder to hit but after cca 20 runs I got unhandled status 0x00800000. Not a regression.

*** This bug has been marked as a duplicate of bug 71659 ***