Bug 58272

Created attachment 71477 [details]
errors in kern log before dma changes

Hmm I see using the word  a t t a c h m e n t does strange things  - 1 and 2 are not mine.

It seems that ttm_mem_evict_first is called way more often in a nasted fashion than is healthy there. Could you resolve the ttm_mem_evict_first address where it ends up calling itself back to a specific line?

It looks nasty though, could you also dump mem_type for each time it calls ttm_mem_evict_first?

Do any of your local patches touch radeon_evict_flags or radeon_ttm_placement_from_domain? I don't see why it would recurse so deeply otherwise.

A full public git tree to reproduce the problem and seeing what patches are applied would also be nice.

Created attachment 71507 [details] [review]
fix

Should be fixed with this patch.

(In reply to comment #6)
> Created attachment 71507 [details] [review] [review]
> fix
> 
> Should be fixed with this patch.

Probably :-)

It seems that current drm-next head + fix has a different issue which makes etqw die quite quickly.

drm-next reset onto 

drm/ttm: remove no_wait_reserve, v3 + the fix 

is now stable with etqw.

The head issue is -

EE r600_texture.c:697 r600_texture_transfer_map - failed to create temporary texture to hold untiled copy
Mesa: User error: GL_OUT_OF_MEMORY in glTexSubImage
radeon: The kernel rejected CS, see dmesg for more information.
double fault: 'Segmentation fault', bailing out

in dmesg - 

[drm:radeon_cs_ioctl] *ERROR* Failed to parse relocation -12!
[drm:radeon_cs_ioctl] *ERROR* Failed to parse relocation -12!
[drm:radeon_cs_ioctl] *ERROR* Failed to parse relocation -12!
[drm:radeon_cs_ioctl] *ERROR* Failed to parse relocation -12!
[drm:radeon_gem_object_create] *ERROR* Failed to allocate GEM object (8192, 2, 4096, -12)
[drm:radeon_cs_ioctl] *ERROR* Failed to parse relocation -12!
etqw.x86[2478]: segfault at 0 ip af5142ad sp bff8b310 error 4 in gamex86.so[af23f000+948000]

Could you please run a git bisection to see where that error has been introduced, then?

Created attachment 71530 [details]
gpu lock + oops on use async dma for ttm buffer moves on 6xx-SI

(In reply to comment #8)
> Could you please run a git bisection to see where that error has been
> introduced, then?

It seems that drm/radeon: use async dma for ttm buffer moves on 6xx-SI is the first non working, but it gives a different fail from head. Log attached.

A patch referencing this bug report has been merged in Linux v3.8-rc1:

commit dd54fee7d440c4a9756cce2c24a50c15e4c17ccb
Author: Dave Airlie <airlied@redhat.com>
Date:   Fri Dec 14 21:04:46 2012 +1000

    radeon: fix regression with eviction since evict caching changes

Make sure your kernel has this patch:
http://git.kernel.org/?p=linux/kernel/git/torvalds/linux.git;a=commitdiff;h=0953e76e91f4b6206cef50bd680696dc6bf1ef99

(In reply to comment #12)
> Make sure your kernel has this patch:
> http://git.kernel.org/?p=linux/kernel/git/torvalds/linux.git;a=commitdiff;
> h=0953e76e91f4b6206cef50bd680696dc6bf1ef99

I tested drm-next head when that went in and got the same results.

I've just rebuilt it to be sure and with etqw I get a segfault after about 10 secs and in dmesg -

[drm:radeon_cs_ioctl] *ERROR* Failed to parse relocation -12!

I've also managed to reproduce the GPU lock + oops I reported earlier - this time with nexuiz on current drm-next head.

I am not getting ttm errors any more so I guess this bug should be closed?

(In reply to comment #13)
> (In reply to comment #12)
> > Make sure your kernel has this patch:
> > http://git.kernel.org/?p=linux/kernel/git/torvalds/linux.git;a=commitdiff;
> > h=0953e76e91f4b6206cef50bd680696dc6bf1ef99
> 
> I tested drm-next head when that went in and got the same results.
> 
> I've just rebuilt it to be sure and with etqw I get a segfault after about
> 10 secs and in dmesg -
> 
> [drm:radeon_cs_ioctl] *ERROR* Failed to parse relocation -12!
> 
> I've also managed to reproduce the GPU lock + oops I reported earlier - this
> time with nexuiz on current drm-next head.
> 
> I am not getting ttm errors any more so I guess this bug should be closed?

FWIW I tried current drm-next + patch -

0003-drm-radeon-fix-dma-copy-on-r6xx-r7xx-evergen-ni-si-g.patch

And I still fail with etqw after about 10 secs, but do get more info.

radeon: The kernel rejected CS, see dmesg for more information.
radeon: The kernel rejected CS, see dmesg for more information.
radeon: The kernel rejected CS, see dmesg for more information.
radeon: Failed to allocate a buffer:
radeon:    size      : 7168 bytes
radeon:    alignment : 256 bytes
radeon:    domains   : 2
EE r600_texture.c:697 r600_texture_transfer_map - failed to create temporary texture to hold untiled copy
Mesa: User error: GL_OUT_OF_MEMORY in glTexSubImage
radeon: The kernel rejected CS, see dmesg for more information.
radeon: The kernel rejected CS, see dmesg for more information.
radeon: The kernel rejected CS, see dmesg for more information.
radeon: The kernel rejected CS, see dmesg for more information.
radeon: The kernel rejected CS, see dmesg for more information.
radeon: The kernel rejected CS, see dmesg for more information.
radeon: The kernel rejected CS, see dmesg for more information.
double fault: 'Segmentation fault', bailing out
shutdown terminal support
/home/andy/bin/etqw: line 1:  2472 Segmentation fault      /usr/local/games/etqw/etqw

dmesg -

[drm:radeon_cs_ioctl] *ERROR* Failed to parse relocation -12!
[drm:radeon_cs_ioctl] *ERROR* Failed to parse relocation -12!
[drm:radeon_cs_ioctl] *ERROR* Failed to parse relocation -12!
[drm:radeon_gem_object_create] *ERROR* Failed to allocate GEM object (8192, 2, 4096, -12)
[drm:radeon_cs_ioctl] *ERROR* Failed to parse relocation -12!
[drm:radeon_cs_ioctl] *ERROR* Failed to parse relocation -12!
[drm:radeon_cs_ioctl] *ERROR* Failed to parse relocation -12!
[drm:radeon_cs_ioctl] *ERROR* Failed to parse relocation -12!
[drm:radeon_cs_ioctl] *ERROR* Failed to parse relocation -12!
[drm:radeon_cs_ioctl] *ERROR* Failed to parse relocation -12!
[drm:radeon_cs_ioctl] *ERROR* Failed to parse relocation -12!
etqw.x86[2472]: segfault at 0 ip af5292ad sp bfbe3250 error 4 in gamex86.so[af254000+948000]

Current drm-fixes is working for me now.

The remaining etqw issue was fixed by -

Revert "drm/radeon: do not move bo to different placement at each cs"