amdgpu crashes on OOM

amdgpu crashes on OOM

[Alex Xu (Hello71)]

Hi,

I frequently encounter OOM on my system, mostly due to my own fault. 
Recently, I noticed that not only does a swap storm happen and OOM 
killer gets invoked, but the graphics output freezes permanently. 
Checking the kernel messages, I see:

kworker/u24:4: page allocation failure: order:5, mode:0x40dc0(GFP_KERNEL|__GFP_COMP|__GFP_ZERO), nodemask=(null)
CPU: 6 PID: 279469 Comm: kworker/u24:4 Tainted: G        W         5.9.0-14732-g20b1adb60cf6 #2
Hardware name: To Be Filled By O.E.M. To Be Filled By O.E.M./B450 Pro4, BIOS P4.20 06/18/2020
Workqueue: events_unbound commit_work
Call Trace:
 ? dump_stack+0x57/0x6a
 ? warn_alloc.cold+0x69/0xcd
 ? __alloc_pages_direct_compact+0xfb/0x116
 ? __alloc_pages_slowpath.constprop.0+0x9c2/0xc14
 ? __alloc_pages_nodemask+0x143/0x167
 ? kmalloc_order+0x24/0x64
 ? dc_create_state+0x1a/0x4d
 ? amdgpu_dm_atomic_commit_tail+0x1b19/0x227d

followed by:

WARNING: CPU: 6 PID: 279469 at drivers/gpu/drm/amd/amdgpu/../display/amdgpu_dm/amdgpu_dm.c:7511 amdgpu_dm_atomic_commit_tail+0x217c/0x227d

followed by:

BUG: unable to handle page fault for address: 0000000000012480
#PF: supervisor read access in kernel mode
#PF: error_code(0x0000) - not-present page
[ ... ]
RIP: 0010:dc_resource_state_copy_construct+0x10/0x455
[ ... ]
Call Trace:
 ? amdgpu_dm_atomic_commit_tail+0x2193/0x227

This area of code is quite odd:

dc_state_temp = dc_create_state(dm->dc);
ASSERT(dc_state_temp);
dc_state = dc_state_temp;
dc_resource_state_copy_construct_current(dm->dc, dc_state);

This ASSERT macro is misleading: unless CONFIG_DEBUG_KERNEL_DC is set, 
it is actually WARN_ON_ONCE(!(expr)). Therefore, this code fails to 
allocate memory (causing a warning to be printed), prints another 
warning that it failed, then proceeds to immediately dereference it, 
crashing the thread (and the kernel if panic_on_oops is set).

While I am not by any means a graphics or kernel expert, it seems to me 
like there should be a better solution than crashing. If nothing else, 
the OOM killer should be invoked and the operation retried. We may lose 
some frames or see some corruption, but that's far better than totally 
breaking.

Thanks,
Alex.

Re: amdgpu crashes on OOM

[Michel Dänzer]

On 2020-10-26 5:29 a.m., Alex Xu (Hello71) wrote:
> Hi,
> 
> I frequently encounter OOM on my system, mostly due to my own fault.
> Recently, I noticed that not only does a swap storm happen and OOM
> killer gets invoked, but the graphics output freezes permanently.
> Checking the kernel messages, I see:
> 
> kworker/u24:4: page allocation failure: order:5, mode:0x40dc0(GFP_KERNEL|__GFP_COMP|__GFP_ZERO), nodemask=(null)
> CPU: 6 PID: 279469 Comm: kworker/u24:4 Tainted: G        W         5.9.0-14732-g20b1adb60cf6 #2
> Hardware name: To Be Filled By O.E.M. To Be Filled By O.E.M./B450 Pro4, BIOS P4.20 06/18/2020
> Workqueue: events_unbound commit_work
> Call Trace:
>   ? dump_stack+0x57/0x6a
>   ? warn_alloc.cold+0x69/0xcd
>   ? __alloc_pages_direct_compact+0xfb/0x116
>   ? __alloc_pages_slowpath.constprop.0+0x9c2/0xc14
>   ? __alloc_pages_nodemask+0x143/0x167
>   ? kmalloc_order+0x24/0x64
>   ? dc_create_state+0x1a/0x4d
>   ? amdgpu_dm_atomic_commit_tail+0x1b19/0x227d

Looks like dc_create_state should use kvzalloc instead of kzalloc 
(dc_state_free already uses kvfree).

order:5 means it's trying to allocate 32 physically contiguous pages, 
which can be hard to fulfill even with lower memory pressure.


-- 
Earthling Michel Dänzer               |               https://redhat.com
Libre software enthusiast             |             Mesa and X developer

RE: amdgpu crashes on OOM

[Deucher, Alexander]

[AMD Public Use]

> -----Original Message-----
> From: Michel Dänzer <michel@daenzer.net>
> Sent: Monday, October 26, 2020 7:04 AM
> To: Alex Xu (Hello71) <alex_y_xu@yahoo.ca>; Kazlauskas, Nicholas
> <Nicholas.Kazlauskas@amd.com>; Deucher, Alexander
> <Alexander.Deucher@amd.com>; Wentland, Harry
> <Harry.Wentland@amd.com>; Li, Sun peng (Leo) <Sunpeng.Li@amd.com>;
> amd-gfx@lists.freedesktop.org
> Cc: linux-kernel@vger.kernel.org
> Subject: Re: amdgpu crashes on OOM
> 
> On 2020-10-26 5:29 a.m., Alex Xu (Hello71) wrote:
> > Hi,
> >
> > I frequently encounter OOM on my system, mostly due to my own fault.
> > Recently, I noticed that not only does a swap storm happen and OOM
> > killer gets invoked, but the graphics output freezes permanently.
> > Checking the kernel messages, I see:
> >
> > kworker/u24:4: page allocation failure: order:5,
> mode:0x40dc0(GFP_KERNEL|__GFP_COMP|__GFP_ZERO),
> nodemask=(null)
> > CPU: 6 PID: 279469 Comm: kworker/u24:4 Tainted: G        W         5.9.0-14732-
> g20b1adb60cf6 #2
> > Hardware name: To Be Filled By O.E.M. To Be Filled By O.E.M./B450
> > Pro4, BIOS P4.20 06/18/2020
> > Workqueue: events_unbound commit_work
> > Call Trace:
> >   ? dump_stack+0x57/0x6a
> >   ? warn_alloc.cold+0x69/0xcd
> >   ? __alloc_pages_direct_compact+0xfb/0x116
> >   ? __alloc_pages_slowpath.constprop.0+0x9c2/0xc14
> >   ? __alloc_pages_nodemask+0x143/0x167
> >   ? kmalloc_order+0x24/0x64
> >   ? dc_create_state+0x1a/0x4d
> >   ? amdgpu_dm_atomic_commit_tail+0x1b19/0x227d
> 
> Looks like dc_create_state should use kvzalloc instead of kzalloc
> (dc_state_free already uses kvfree).
> 
> order:5 means it's trying to allocate 32 physically contiguous pages, which can
> be hard to fulfill even with lower memory pressure.
> 

It was using kvzalloc, but was accidently dropped when that code was refactored.  I just sent a patch to fix it.

Alex

> 
> --
> Earthling Michel Dänzer               |
> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fredh
> at.com%2F&amp;data=04%7C01%7Calexander.deucher%40amd.com%7Cc60
> 56551dd4d423bdc0508d8799ed189%7C3dd8961fe4884e608e11a82d994e183d
> %7C0%7C0%7C637393070333648663%7CUnknown%7CTWFpbGZsb3d8eyJWIj
> oiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1
> 000&amp;sdata=a7Lpu04KnpsFQpCO7y5WOLJSMPpA%2Be1s%2FufgYTDHs2k
> %3D&amp;reserved=0
> Libre software enthusiast             |             Mesa and X developer

RE: amdgpu crashes on OOM

[Alex Xu (Hello71)]

Excerpts from Deucher, Alexander's message of October 26, 2020 10:34 am:
> It was using kvzalloc, but was accidently dropped when that code was refactored.  I just sent a patch to fix it.

Ah, that explains why I wasn't seeing it before. I was only looking at 
changes in amdgpu_dm_atomic_commit_tail, not dc_create_state.

Thanks,
Alex.