[PATCH] drm/kfd: fix a system crash issue during GPU recovery

[Felix Kuehling]

On 2020-09-01 11:21 a.m., Li, Dennis wrote:
> [AMD Official Use Only - Internal Distribution Only]
>
> Hi, Felix,
>       If GPU hang, execute_queues_cpsch will fail to unmap or map queues and then create_queue_cpsch will return error. If pqm_create_queue find create_queue_cpsch failed, it will call uninit_queue to free queue object. However this queue object has been added in to qpd->queues_list in the old code.

Right, that's a problem. I think the intention here is to keep going 
because a failure to execute the runlist affects not just the queue that 
was just created, but all queues in all processes.

The failure to execute_queues should probably not be reported to the 
caller of create_queue, because the queue was already created, and the 
problem with execute_queues has a much bigger scope than this one 
caller. So I think the correct solution is to ignore the return value 
from execute_queues.

As a follow up, we should probably handle all the error scenarios inside 
execute_queues and make it a void function. Failure to unmap queues 
already triggers a GPU reset, so nothing new needs to be done for that. 
But we need to add handling of failures to map queues. It doesn't 
require a GPU reset, because the problem is in the kernel (e.g. out of 
memory), not the GPU. The best we can do is report this asynchronously 
as a GPU hang to all KFD processes, so they know the GPU is no longer 
going to work for them.

Regards,
   Felix

>
> Best Regards
> Dennis Li
>
> -----Original Message-----
> From: Kuehling, Felix <Felix.Kuehling at amd.com>
> Sent: Tuesday, September 1, 2020 9:26 PM
> To: Li, Dennis <Dennis.Li at amd.com>; amd-gfx at lists.freedesktop.org; Deucher, Alexander <Alexander.Deucher at amd.com>; Zhang, Hawking <Hawking.Zhang at amd.com>; Koenig, Christian <Christian.Koenig at amd.com>
> Subject: Re: [PATCH] drm/kfd: fix a system crash issue during GPU recovery
>
> I'm not sure how the bug you're fixing is caused, but your fix is clearly in the wrong place.
>
> A queue being disabled is not the same thing as a queue being destroyed.
> Queues can be disabled for legitimate reasons, but they still should exist and be in the qpd->queues_list.
>
> If a destroyed queue is left on the qpd->queues_list, that would be a problem. Can you point out where such a thing is happening?
>
> Thanks,
>    Felix
>
>
> Am 2020-08-31 um 9:36 p.m. schrieb Dennis Li:
>> The crash log as the below:
>>
>> [Thu Aug 20 23:18:14 2020] general protection fault: 0000 [#1] SMP NOPTI
>> [Thu Aug 20 23:18:14 2020] CPU: 152 PID: 1837 Comm: kworker/152:1 Tainted: G           OE     5.4.0-42-generic #46~18.04.1-Ubuntu
>> [Thu Aug 20 23:18:14 2020] Hardware name: GIGABYTE
>> G482-Z53-YF/MZ52-G40-00, BIOS R12 05/13/2020 [Thu Aug 20 23:18:14
>> 2020] Workqueue: events amdgpu_ras_do_recovery [amdgpu] [Thu Aug 20
>> 23:18:14 2020] RIP: 0010:evict_process_queues_cpsch+0xc9/0x130
>> [amdgpu] [Thu Aug 20 23:18:14 2020] Code: 49 8d 4d 10 48 39 c8 75 21
>> eb 44 83 fa 03 74 36 80 78 72 00 74 0c 83 ab 68 01 00 00 01 41 c6 45
>> 41 00 48 8b 00 48 39 c8 74 25 <80> 78 70 00 c6 40 6d 01 74 ee 8b 50 28
>> c6 40 70 00 83 ab 60 01 00 [Thu Aug 20 23:18:14 2020] RSP:
>> 0018:ffffb29b52f6fc90 EFLAGS: 00010213 [Thu Aug 20 23:18:14 2020] RAX:
>> 1c884edb0a118914 RBX: ffff8a0d45ff3c00 RCX: ffff8a2d83e41038 [Thu Aug
>> 20 23:18:14 2020] RDX: 0000000000000000 RSI: 0000000000000082 RDI:
>> ffff8a0e2e4178c0 [Thu Aug 20 23:18:14 2020] RBP: ffffb29b52f6fcb0 R08:
>> 0000000000001b64 R09: 0000000000000004 [Thu Aug 20 23:18:14 2020] R10:
>> ffffb29b52f6fb78 R11: 0000000000000001 R12: ffff8a0d45ff3d28 [Thu Aug 20 23:18:14 2020] R13: ffff8a2d83e41028 R14: 0000000000000000 R15: 0000000000000000 [Thu Aug 20 23:18:14 2020] FS:  0000000000000000(0000) GS:ffff8a0e2e400000(0000) knlGS:0000000000000000 [Thu Aug 20 23:18:14 2020] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [Thu Aug 20 23:18:14 2020] CR2: 000055c783c0e6a8 CR3: 00000034a1284000 CR4: 0000000000340ee0 [Thu Aug 20 23:18:14 2020] Call Trace:
>> [Thu Aug 20 23:18:14 2020]  kfd_process_evict_queues+0x43/0xd0
>> [amdgpu] [Thu Aug 20 23:18:14 2020]
>> kfd_suspend_all_processes+0x60/0xf0 [amdgpu] [Thu Aug 20 23:18:14
>> 2020]  kgd2kfd_suspend.part.7+0x43/0x50 [amdgpu] [Thu Aug 20 23:18:14
>> 2020]  kgd2kfd_pre_reset+0x46/0x60 [amdgpu] [Thu Aug 20 23:18:14 2020]
>> amdgpu_amdkfd_pre_reset+0x1a/0x20 [amdgpu] [Thu Aug 20 23:18:14 2020]
>> amdgpu_device_gpu_recover+0x377/0xf90 [amdgpu] [Thu Aug 20 23:18:14
>> 2020]  ? amdgpu_ras_error_query+0x1b8/0x2a0 [amdgpu] [Thu Aug 20
>> 23:18:14 2020]  amdgpu_ras_do_recovery+0x159/0x190 [amdgpu] [Thu Aug
>> 20 23:18:14 2020]  process_one_work+0x20f/0x400 [Thu Aug 20 23:18:14
>> 2020]  worker_thread+0x34/0x410
>>
>> When GPU hang, user process will fail to create a compute queue whose
>> struct object will be freed later, but driver wrongly add this queue
>> to queue list of the proccess. And then kfd_process_evict_queues will
>> access a freed memory, which cause a system crash.
>>
>> Signed-off-by: Dennis Li <Dennis.Li at amd.com>
>>
>> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
>> b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
>> index 560adc57a050..d5e6b07ffb27 100644
>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
>> @@ -1296,16 +1296,18 @@ static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q,
>>   	mqd_mgr->init_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj,
>>   				&q->gart_mqd_addr, &q->properties);
>>   
>> -	list_add(&q->list, &qpd->queues_list);
>> -	qpd->queue_count++;
>> -
>>   	if (q->properties.is_active) {
>>   		increment_queue_count(dqm, q->properties.type);
>>   
>>   		retval = execute_queues_cpsch(dqm,
>>   				KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
>> +		if (retval)
>> +			goto out_execute_cpsch;
>>   	}
>>   
>> +	list_add(&q->list, &qpd->queues_list);
>> +	qpd->queue_count++;
>> +
>>   	/*
>>   	 * Unconditionally increment this counter, regardless of the queue's
>>   	 * type or whether the queue is active.
>> @@ -1318,6 +1320,9 @@ static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q,
>>   	dqm_unlock(dqm);
>>   	return retval;
>>   
>> +out_execute_cpsch:
>> +	decrement_queue_count(dqm, q->properties.type);
>> +	dqm_unlock(dqm);
>>   out_deallocate_doorbell:
>>   	deallocate_doorbell(qpd, q);
>>   out_deallocate_sdma_queue: