| Bug ID | 97594 |
|---|---|
| Summary | [amdgpu SI] "drm/amd/amdgpu: Add GRBM lock to various SI= functions" breaks amdgpu support for SI |
| Product | DRI |
| Version | unspecified |
| Hardware | Other |
| OS | All |
| Status | NEW |
| Severity | normal |
| Priority | medium |
| Component | DRM/AMDgpu |
| Assignee | dri-devel@lists.freedesktop.org |
| Reporter | arek.rusi@gmail.com |
Created attachment 12=
6204 [details]
dmesg for first bad commit
With this commit displays off during loading amdgpu, sysrq don't work, boot=
is
stopped.=20
When i reverted it, amdgpu works as always.
dmesg is silent
Created attachment 1262=
05 [details]
commit: deca1d1 reverted
sorry for lack of description It's Alex's drm-next-4.9-wip branch. First bad commit:=20 commit deca1d1f16eebfa0d070eed50a221e01cf716ee0 author Tom St Denis <tom.std= enis@amd.com> drm/amd/amdgpu: Add GRBM lock to various SI functions Add missing lock around SE/SH/INSTANCE selections.
(In reply to Arek Ru=C5=9Bniak from comment #0) > Created attachment 126204 [details] > dmesg for first bad commit >=20 > With this commit displays off during loading amdgpu, sysrq don't work,= boot > is stopped.=20 > When i reverted it, amdgpu works as always. >=20 > dmesg is silent Same behavior but on PREEMPT kernel even without mentioned patch. Just to clarify that it could be more general problem than locking in this particul= ar place.
Could you share config and patchset for your preempt kernel. I= f I will have some time I could build it. I'm just curious. I know nothing about what preempt is. I think I should rather call my kerne= l - "generic" but "uname -a" gives me: Linux darkfruit 4.8.0-rc1-20160904 #10 SMP PREEMPT Tue Sep 6 23:01:48 CEST = 2016 x86_64 GNU/Linux So I'm confused about it
| What | Removed | Added |
|---|---|---|
| CC | tom.stdenis@amd.com |
The problem is a double lock. The grbm_idx_mutex lock is take= n in gfx_v6_0_get_cu_info() which then calls gfx_v6_0_get_cu_active_bitmap() whi= ch also takes the lock. The fix is to remove the lock and the select from the parent function. I'll attach a patch momentarily.
The problem is a double lock. The grbm_idx_mutex lock is take= n in gfx_v6_0_get_cu_info() which then calls gfx_v6_0_get_cu_active_bitmap() whi= ch also takes the lock. The fix is to remove the lock and the select from the parent function. I'll attach a patch momentarily.
Created attachment 1= 26274 [details] [review] Patch to remove double lock This removes the double lock. It's actually patch #9 in my yet to be RB'ed series of SI cleanups ...
| What | Removed | Added |
|---|---|---|
| Status | NEW | RESOLVED |
| Resolution | --- | FIXED |