add BSW in this bug.

Chris, did you mean this issue is platform interrelated for you removed the platform 'BDW'?

It's a bug in the capture code that is not specific to any architecture.

(In reply to Chris Wilson from comment #3)
> It's a bug in the capture code that is not specific to any architecture.

(The issue is magnified by the partial seqno/request conversion.)

*** Bug 88821 has been marked as a duplicate of this bug. ***

*** Bug 89441 has been marked as a duplicate of this bug. ***

Does development team agree this as highest priority? If so can we move on?

Lost track here ... have we merged the patches Chris?

No, it is something that I addressed in the conversion to requests but has been overlooked.

Reducing bug priority after a discussion with Chris. Main points are

 - the bug is not a regression, it has been in the code base since the introcution of lockless error capture;
 - there is no user sighting of the bug;
 - the blocked test case (gem_evict_everything/swapping-hang) tests for an extreme corner-case.


Also, according to Chris, the for the solution "we need a couple of spinlocks to serialize bo retirement vs error capture, but we need to avoid creating deadlocks, and that is the tricky part."

Created attachment 117649 [details]
HSW-ULT_dmesg.txt

Hi, this issue also occurs with the latest configuration for HSW-ULT

-- Hardware --
Platform: Intel NUC D54250WYK
Processo: Intel(R) Core(TM) i5-4250U CPU @ 1.30GHz
-- Software --
Linux distribution: Ubuntu 14.04.02 LTS 64Bits
BIOS: WYLPT10H.86A.0021.2013.1017.1606



Test Environment:
````````````````````````````````````
Kernel: tag drm-intel-testing-2015-07-31 (4.2-rc4) from git://anongit.freedesktop.org/drm-intel
Mesa: mesa-10.6.3 from http://cgit.freedesktop.org/mesa/mesa/
Xf86_video_intel: 2.99.917 from http://cgit.freedesktop.org/xorg/driver/xf86-video-intel/
Libdrm: libdrm-2.4.62 from http://cgit.freedesktop.org/mesa/drm/
Cairo: 1.14.2 from http://cgit.freedesktop.org/cairo
libva: libva-1.6.0 from http://cgit.freedesktop.org/libva/
intel-driver: 1.6.0. from http://cgit.freedesktop.org/vaapi/intel-driver
xorg: 1.17.99 installed with script git_xorg.sh
Xserver: xorg-server-1.17.2 from http://cgit.freedesktop.org/xorg/xserver
Intel-gpu-tools: 1.11 from http://cgit.freedesktop.org/xorg/app/intel-gpu-tools/


Notes : It often causes system hang. Fail rate : 4/5,  and sometimes causes dmesg warning
	Attached HSW-ULT_dmesg.txt


If needed more information or you have any doubt do not hesitate to contact me

Created attachment 117672 [details]
BDW-U dmesg log

Bug scrub:
Probably fixed, can you confirm?

No. Error capture still dereferences requests without any serialisation with the freeing of said requests.

Created attachment 118888 [details]
BDW dmesg log

Bug Scrub:

Tested again on BDW using kernel 4.3.0 and got an error as well, find attached the dmesg log and find below the Environment I used

````````````````````````````````````
Kernel:4.3.0-rc4 drm-intel-testing-2015-10-10
Mesa: mesa-11.0.2 
Xf86_video_intel: 2.99.917 
Libdrm: libdrm-2.4.65
Cairo: 1.14.2 
libva: libva-1.6.1 
intel-driver: 1.6.1
xorg: 1.17.99 installed with script git_xorg.sh
Xserver: xorg-server-1.17.2 
Intel-gpu-tools: 1.12

Bug scrub,
Assigned to Kimmo

http://patchwork.freedesktop.org/patch/70010/

(In reply to Chris Wilson from comment #17)
> http://patchwork.freedesktop.org/patch/70010/

Can anybody please confirm whether the patch above solves the problem or at least reduces the failure rate?

Thanks,
Paulo

Jairo, please re-test with the patch and confirm if it is still occuring.

Seems that the patch is not valid for drm-intel-next-2016-05-08-2069-gf1eaed1.. equivalent for drm-intel-testing-05-21-2016. 

The file i915_gpu_error.c is not taking the patches.

Hunk #3 FAILED at 1290.
1 out of 3 hunks FAILED -- saving rejects to file drivers/gpu/drm/i915/i915_gpu_error.c.rej
(04:05 AM) [gfx@gfx-ThinkCentre-M600] [drm-intel]$ : nano drivers/gpu/drm/i915/i915_gpu_error.c.rej
  GNU nano 2.5.3   File: drivers/gpu/drm/i915/i915_gpu_error.c.rej

--- drivers/gpu/drm/i915/i915_gpu_error.c
+++ drivers/gpu/drm/i915/i915_gpu_error.c
@@ -1290,9 +1269,19 @@ void i915_capture_error_state(struct drm_device *dev, bo$
        }

        kref_init(&error->ref);
-       error->i915 = dev_priv;

-       stop_machine(capture, error, NULL);
+       i915_capture_gen_state(dev_priv, error);
+       i915_capture_reg_state(dev_priv, error);
+       i915_gem_record_fences(dev, error);
+       i915_gem_record_rings(dev, error);
+
+       i915_capture_active_buffers(dev_priv, error);
+       i915_capture_pinned_buffers(dev_priv, error);
+
+       do_gettimeofday(&error->time);
+

(In reply to Chris Wilson from comment #17)
> http://patchwork.freedesktop.org/patch/70010/

HI Chris, this patch we could not apply im the latest kernels 4.7.0-rc7, could you do a double check please?

Well, we are getting closer it is only at about patch 90 in the queue now. The patch in situ is https://cgit.freedesktop.org/~ickle/linux-2.6/commit/?h=tasklet&id=c9a8be989704c323a87c2fd661b3a65815daa938

This test is now being skipped due to "lack of memory", I tested in BXT and  SKL using the following Kernel:


===================================================================
commit 57de27e40b9741c17c6749a366e891faf8b22fcb
Author: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Date:   Mon Aug 29 17:38:46 2016 +0200

    drm-intel-nightly: 2016y-08m-29d-15h-38m-26s UTC integration manifest
===================================================================

I am getting the following message 

IGT-Version: 1.15-g572a770 (x86_64) (Linux: 4.8.0-rc4drm-intel-nighly-ww35-commi                                   64)
Test requirement not met in function intel_require_memory, file intel_os.c:289:
Test requirement: __intel_check_memory(count, size, mode, &required, &total)
Estimated that we need 201,326,592 objects and 201,424,896 MiB for the test, but                                   89 MiB available (RAM) and a maximum of 1,611,544 objects


Notice the " estimated " memory required is an abnormal amount of memory.

(In reply to Jairo Miramontes from comment #23)
> I am getting the following message 
> 
> IGT-Version: 1.15-g572a770 (x86_64) (Linux:
> 4.8.0-rc4drm-intel-nighly-ww35-commi                                   64)
> Test requirement not met in function intel_require_memory, file
> intel_os.c:289:
> Test requirement: __intel_check_memory(count, size, mode, &required, &total)
> Estimated that we need 201,326,592 objects and 201,424,896 MiB for the test,
> but                                   89 MiB available (RAM) and a maximum
> of 1,611,544 objects
> 
> 
> Notice the " estimated " memory required is an abnormal amount of memory.

But accurate. That test is irrelevant regarding this bug. The bug is a race condition in our error capture code that only depends upon running the error capture whilst the driver is active.

commit 9f267eb8d2ea0a87f694da3f236067335e8cb7b9
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date:   Wed Oct 12 10:05:19 2016 +0100

    drm/i915: Stop the machine whilst capturing the GPU crash dump