Re: [PATCH 1/2] drm/amd/display: Protect dml2_create()/dml2_copy()/dml2_create_copy()

[Aurabindo Pillai <aurabindo.pillai@amd.com>]

On 2025-03-26 21:40, Huacai Chen wrote:
> Hi, Alex,
> 
> On Thu, Mar 27, 2025 at 8:10 AM Alex Hung <alex.hung@amd.com> wrote:
>>
>> The following error messages showed up on an APU and a dGPU during testing.
>>
>> <3> [100.231411] BUG: sleeping function called from invalid context at
>> include/linux/sched/mm.h:321
>> <3> [100.231414] in_atomic(): 1, irqs_disabled(): 0, non_block: 0, pid:
>> 1711, name: kms_color
>> <3> [100.231416] preempt_count: 2, expected: 0
>> <3> [100.231417] RCU nest depth: 0, expected: 0
>> <3> [100.231418] Preemption disabled at:
>> <3> [100.231419] [<ffffffffc0c2843b>] dc_fpu_begin+0x2b/0xc0 [amdgpu]
>> <4> [100.231626] CPU: 4 UID: 0 PID: 1711 Comm: kms_color Tainted: G
>>     W          6.12.0+ #1
>> <4> [100.231629] Tainted: [W]=WARN
>> <4> [100.231631] Call Trace:
>> <4> [100.231632]  <TASK>
>> <4> [100.231633]  dump_stack_lvl+0x5b/0x70
>> <4> [100.231638]  dump_stack+0x10/0x20
>> <4> [100.231639]  __might_resched+0x170/0x1d0
>> <4> [100.231643]  __might_sleep+0x44/0x70
>> <4> [100.231645]  __alloc_pages_noprof+0x22f/0x370
>> <4> [100.231649]  ___kmalloc_large_node+0x95/0x150
>> <4> [100.231651]  ? preempt_count_add+0x4e/0xc0
>> <4> [100.231653]  __kmalloc_large_noprof+0x1d/0xb0
>> <4> [100.231655]  dml2_create_copy+0x27/0x60 [amdgpu]
>> <4> [100.231827]  dc_state_create_copy+0x7e/0x170 [amdgpu]
>> <4> [100.231995]  update_planes_and_stream_state+0x23c/0x600 [amdgpu]
>> <4> [100.232189]  update_planes_and_stream_v2+0x22b/0x530 [amdgpu]
>> <4> [100.232366]  ? amdgpu_dm_atomic_commit_tail+0x1310/0x4100 [amdgpu]
>> <4> [100.232569]  ? commit_tail+0x96/0x140 [drm_kms_helper]
>> <4> [100.232577]  dc_update_planes_and_stream+0x5b/0xe0 [amdgpu]
>> <4> [100.232730]  amdgpu_dm_atomic_commit_tail+0x1fa7/0x4100 [amdgpu]
>> <4> [100.232908]  ? stack_depot_save_flags+0x2c/0x730
>> <4> [100.232915]  ? wait_for_completion_timeout+0x1d/0x30
>> <4> [100.232917]  commit_tail+0x96/0x140 [drm_kms_helper]
>> <4> [100.232923]  drm_atomic_helper_commit+0x12b/0x150 [drm_kms_helper]
>> <4> [100.232927]  drm_atomic_commit+0xad/0xe0 [drm]
>> <4> [100.232939]  ? __pfx___drm_printfn_info+0x10/0x10 [drm]
>> <4> [100.232956]  drm_atomic_helper_set_config+0x80/0xc0 [drm_kms_helper]
>> <4> [100.232961]  drm_mode_setcrtc+0x22e/0x910 [drm]
>> <4> [100.232975]  ? kfree+0x18f/0x350
>> <4> [100.232977]  ? __pfx_drm_mode_setcrtc+0x10/0x10 [drm]
>> <4> [100.232987]  drm_ioctl_kernel+0xa7/0x100 [drm]
>> <4> [100.233004]  drm_ioctl+0x29d/0x500 [drm]
>> <4> [100.233015]  ? __pfx_drm_mode_setcrtc+0x10/0x10 [drm]
>> <4> [100.233026]  ? _raw_spin_unlock_irqrestore+0x1f/0x40
>> <4> [100.233029]  amdgpu_drm_ioctl+0x4b/0x80 [amdgpu]
>> <4> [100.233131]  __x64_sys_ioctl+0x92/0xd0
>> <4> [100.233133]  x64_sys_call+0x1205/0x20d0
>> <4> [100.233136]  do_syscall_64+0x50/0x110
>> <4> [100.233138]  entry_SYSCALL_64_after_hwframe+0x76/0x7e
>> <4> [100.233142] RIP: 0033:0x7fb21e71a94f
>> <4> [100.233144] Code: 00 48 89 44 24 18 31 c0 48 8d 44 24 60 c7 04 24
>> 10 00 00 00 48 89 44 24 08 48 8d 44 24 20 48 89 44 24 10 b8 10 00 00 00
>> 0f 05 <41> 89 c0 3d 00 f0 ff ff 77 1f 48 8b 44 24 18 64 48 2b 04 25 28 00
>> <4> [100.233145] RSP: 002b:00007ffdd9a52e50 EFLAGS: 00000246 ORIG_RAX:
>> 0000000000000010
>> <4> [100.233148] RAX: ffffffffffffffda RBX: 00007ffdd9a52ee0 RCX:
>> 00007fb21e71a94f
>> <4> [100.233149] RDX: 00007ffdd9a52ee0 RSI: 00000000c06864a2 RDI:
>> 0000000000000005
>> <4> [100.233149] RBP: 00000000c06864a2 R08: 0000000000000000 R09:
>> 00005609537f7b08
>> <4> [100.233150] R10: 0000000000000000 R11: 0000000000000246 R12:
>> 0000000000000000
>> <4> [100.233151] R13: 0000000000000005 R14: 0000000000000000 R15:
>> 00005609537e2848
>> <4> [100.233152]  </TASK>
> This seems caused by dml2_allocate_memory(), to fix this we can only
> protect FPU in DML2, I can do it in the new version, but I want to
> listen Aurabindo's opinion.
> 
>

It looks like dml21_apply_soc_bb_overrides() does have some division on 
double variables. I'm curious why we dont see this on our side. Was this 
seen on x86 or Loongson?

I think your approach is correct. Thanks for taking time to fix this. We 
can add it to weekly testing if you send us a patch.