[PATCH] drm/amdgpu: add vm fault information to devcoredump

[Christian König]

Am 07.03.24 um 09:37 schrieb Khatri, Sunil:
>
> On 3/7/2024 1:47 PM, Christian König wrote:
>> Am 06.03.24 um 19:19 schrieb Sunil Khatri:
>>> Add page fault information to the devcoredump.
>>>
>>> Output of devcoredump:
>>> **** AMDGPU Device Coredump ****
>>> version: 1
>>> kernel: 6.7.0-amd-staging-drm-next
>>> module: amdgpu
>>> time: 29.725011811
>>> process_name: soft_recovery_p PID: 1720
>>>
>>> Ring timed out details
>>> IP Type: 0 Ring Name: gfx_0.0.0
>>>
>>> [gfxhub] Page fault observed for GPU family:143
>>> Faulty page starting at address 0x0000000000000000
>>> Protection fault status register:0x301031
>>>
>>> VRAM is lost due to GPU reset!
>>>
>>> Signed-off-by: Sunil Khatri <sunil.khatri at amd.com>
>>> ---
>>>   drivers/gpu/drm/amd/amdgpu/amdgpu_reset.c | 15 ++++++++++++++-
>>>   drivers/gpu/drm/amd/amdgpu/amdgpu_reset.h |  1 +
>>>   2 files changed, 15 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_reset.c 
>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_reset.c
>>> index 147100c27c2d..d7fea6cdf2f9 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_reset.c
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_reset.c
>>> @@ -203,8 +203,20 @@ amdgpu_devcoredump_read(char *buffer, loff_t 
>>> offset, size_t count,
>>>                  coredump->ring->name);
>>>       }
>>>   +    if (coredump->fault_info.status) {
>>> +        struct amdgpu_vm_fault_info *fault_info = 
>>> &coredump->fault_info;
>>> +
>>> +        drm_printf(&p, "\n[%s] Page fault observed for GPU 
>>> family:%d\n",
>>> +               fault_info->vmhub ? "mmhub" : "gfxhub",
>>> +               coredump->adev->family);
>>> +        drm_printf(&p, "Faulty page starting at address 0x%016llx\n",
>>> +               fault_info->addr);
>>> +        drm_printf(&p, "Protection fault status register:0x%x\n",
>>> +               fault_info->status);
>>> +    }
>>> +
>>>       if (coredump->reset_vram_lost)
>>> -        drm_printf(&p, "VRAM is lost due to GPU reset!\n");
>>> +        drm_printf(&p, "\nVRAM is lost due to GPU reset!\n");
>>>       if (coredump->adev->reset_info.num_regs) {
>>>           drm_printf(&p, "AMDGPU register dumps:\nOffset:     
>>> Value:\n");
>>>   @@ -253,6 +265,7 @@ void amdgpu_coredump(struct amdgpu_device 
>>> *adev, bool vram_lost,
>>>       if (job) {
>>>           s_job = &job->base;
>>>           coredump->ring = to_amdgpu_ring(s_job->sched);
>>> +        coredump->fault_info = job->vm->fault_info;
>>
>> That's illegal. The VM pointer might already be stale at this point.
>>
>> I think you need to add the fault info of the last fault globally in 
>> the VRAM manager or move this to the process info Shashank is working 
>> on.
>> Are you saying that during the reset or otherwise a vm which is part 
>> of this job could have been freed  and we might have a NULL 
>> dereference or invalid reference? Till now based on the resets and 
>> pagefaults that i have created till now using the same app which we 
>> are using for IH overflow i am able to get the valid vm only.
>>
>> Assuming  amdgpu_vm is freed for this job or stale, are you 
>> suggesting to update this information in adev-> vm_manager along with 
>> existing per vm fault_info or only in vm_manager ?

Good question. having it both in the VM as well as the VM manager sounds 
like the simplest option for now.

Regards,
Christian.

>>
>> Regards,
>> Christian.
>>
>>>       }
>>>         coredump->adev = adev;
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_reset.h 
>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_reset.h
>>> index 60522963aaca..3197955264f9 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_reset.h
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_reset.h
>>> @@ -98,6 +98,7 @@ struct amdgpu_coredump_info {
>>>       struct timespec64               reset_time;
>>>       bool                            reset_vram_lost;
>>>       struct amdgpu_ring            *ring;
>>> +    struct amdgpu_vm_fault_info    fault_info;
>>>   };
>>>   #endif
>>