[bdw-u iommu] DMAR error -> GPU hang

Submitted by Stefan Junker

Assigned to Joonas Lahtinen @dolphin

Link to original bug (#89360)

Description

Created attachment 113867
kernel log

The system randomly freezes and doesn't react to anything afterwards. Not even the magic keys can reboot the system.

Processor model is Intel Core-i7 5500U with the integrated GPU.
Kernel version is 4.0.0-rc1, which is required to even get X / gdm working with the system.

I've attached the kernel log messages which shows an instance of this problem.
Please request any information needed and I'll happily provide it.

steveeJ

Attachment 113867, "kernel log":
drm_i915_crash.log

added Community GPU hang feature: GEM platform: BDW priority::high severity::critical + 1 deleted label

Stefan Junker said:

Short version
---
adding 'intel_iommu=igfx_off' helped


Long version
---
I've tried many things to resolve this issue, from kernel reconfiguration to installing mesa, libdrm, intel drivers from latest repository masters, which all didn't help. I reverted back to the most recent releases of the packages.

From an older forum entry somewhere on the webs I found that this could be related to virtualization techologies and memory remapping, so I added the following arguments to my kernel commandline: 'intel_iommu=igfx_off'

Ever since (about 10-15 hours of very active usage) I haven't had a single freeze.

I still think this is not normal behavior, since turning off iommu for the GPU can't be the right or necessary thing to do.

Bobby Powers uploaded an attachment:

Attachment 113998, "bpowers' dmesg":
gpu.hang.1

Bobby Powers uploaded an attachment:

Attachment 113999, "bpowers' GPU crash dump":
gpu.hang.error.1

Bobby Powers said:

I'm also seeing this, on an i7-5600U. Kernel 4.0.0-rc2+, xf86-video-intel from a few days ago (9fb8154), mesa 10.5-rc3.

I turned off VT-d in the BIOS, and haven't seen any issues since. I've attached both my dmesg + GPU crash dump from one of the times this happened.

Bobby Powers said:

some additional info - my BIOS (3rd gen x1 carbon) apparently marks x2apic as broken. I booted a number of times with intremap=no_x2apic_optout on the kernel command line, and saw what steveej mentioned: a hard freeze.

The system did have the foresight to save the dmesg into the EFI pstore. I have those logs if they are useful.

After removing no_x2apic_optout, the kernel "Enabled IRQ remapping in xapic mode", and under xapic some of the time the kernel was able to recover/reset the chip to an ok-enough state that I could save dmesg and grap the GPU dump from /sys/class/drm/card0/error.

Chris Wilson @ickle closed a related bug:

*** Bug 90823 has been marked as a duplicate of this bug. ***

Chris Wilson @ickle closed a related bug:

*** Bug 90091 has been marked as a duplicate of this bug. ***

Chris Wilson @ickle closed a related bug:

*** Bug 91076 has been marked as a duplicate of this bug. ***

Chris Wilson @ickle said:

In bug 91076, the hangs are anything but random. Simply initialising execlists is enough for the GPU to die.

Chris Wilson @ickle closed a related bug:

*** Bug 91152 has been marked as a duplicate of this bug. ***

Chris Wilson @ickle said:

Quick experiment from bug 91152 indicates that this is a problem with or without execlists enabled.

Chris Wilson @ickle closed a related bug:

*** Bug 91458 has been marked as a duplicate of this bug. ***

Maarten Lankhorst said:

I've noticed the same problem here. I also get frequent DMAR errors without hangs during ubuntu's fade to black animation for dpms off.

Matthias Nagel said:

I found the same problem as in comment 4. If I disable VT-d in the BIOS the crashes disappear. But then I get random segmentation faults from GCC if I try to compile QtWebKit (N.b: I have gentoo and compile all packages by myself.) Hence, I have two options

(1) Disable VT-d for daily work such that i915 does not crash
(2) Enable VT-d and only boot into text console mode if I need to compile QtWebKit

My hardware is

Lenovo Thinkpad X1 Carbon (3rd generation, type 20BT)
Processor Intel Core i5-5200U
Memory: 8 GB PC3-12800L
BIOS: Phoenix, ver. 1.09, 7/22/2015

Interestingly, MemTest86 does not report any memory error in either case (with and without VT-d)

Maarten Lankhorst said:

Sounds more like an actual corruption prevented by DMAR then..

Chris Wilson @ickle closed a related bug:

*** Bug 91633 has been marked as a duplicate of this bug. ***

Chris Wilson @ickle closed a related bug:

*** Bug 91764 has been marked as a duplicate of this bug. ***

siflfran@hawo.net said:

I'm not sure if 91764 really is a duplicate of this bug. intel_iommu=igfx_off doesn't help in my case.

klondike said:

In my case this issue (googling for the opcode hanging the GPU lead me to this bug) was solved by disabling the EFI Framebuffer on the kernel configuration.

If the devs want I can open a second bug to request the Intel GFX drivers taking over from early framebuffers (for example EFI or VGA) to prevent my particular issues.