Re: [PATCH v4 2/2] drm/xe/guc: Preserve HW error state in wedged mode 2

["Dong, Zhanjun" <zhanjun.dong@intel.com>]

On 2026-06-29 7:31 p.m., Daniele Ceraolo Spurio wrote:
> 
> 
> On 6/29/2026 7:08 AM, Zhanjun Dong wrote:
>> When HW reports an uncorrectable error, debugging the root cause may
>> require halting the GPU in the error state so scan dumps can be collected
>> and inspected.
>>
>> Use wedged_mode=2 for this debug mode. In this mode, Xe already avoids 
>> GPU
>> reset on hangs, so also skip the engine reset cleanup path for GuC
>> uncorrectable error reports to keep the HW state intact.
>>
>> Signed-off-by: Zhanjun Dong <zhanjun.dong@intel.com>
>> ---
>>   drivers/gpu/drm/xe/xe_guc_submit.c | 6 ++++--
>>   1 file changed, 4 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/xe/xe_guc_submit.c b/drivers/gpu/drm/xe/ 
>> xe_guc_submit.c
>> index 29ccabc4b7a4..05fa4e466461 100644
>> --- a/drivers/gpu/drm/xe/xe_guc_submit.c
>> +++ b/drivers/gpu/drm/xe/xe_guc_submit.c
>> @@ -3024,8 +3024,10 @@ int xe_guc_uncorrectable_error_handler(struct 
>> xe_guc *guc, u32 *msg, u32 len)
>>       trace_xe_guc_uncorrectable_error(q);
>> -    /* Treat the same as engine reset */
>> -    xe_guc_exec_queue_reset_trigger_cleanup(q);
>> +    if (guc_to_xe(guc)->wedged.mode != 
>> XE_WEDGED_MODE_UPON_ANY_HANG_NO_RESET) {
>> +        /* Treat the same as engine reset */
>> +        xe_guc_exec_queue_reset_trigger_cleanup(q);
>> +    }
> 
> This doesn't seem the correct place to do this. On one side, if we skip 
> the reset call here, if the queue it's actually hung on the engine then 
> the fence will eventually timeout and guc_exec_queue_timedout_job() will 
> be called anyway. On the other side, 
> XE_WEDGED_MODE_UPON_ANY_HANG_NO_RESET is a generic flag and 
> xe_guc_exec_queue_reset_trigger_cleanup is also called from the CAT 
> error path, so we should cover that as well. The best way to solve both 
> issues IMO is to add a check in guc_exec_queue_timedout_job to wedge if 
> XE_WEDGED_MODE_UPON_ANY_HANG_NO_RESET is set.

Good point, test shown the existing mode=2 feature already handles this, 
so this patch is out of scope and would be removed in next rev.

Regards,
Zhanjun Dong>
> Daniele
> 
>>       return 0;
>>   }
>