Re: [PATCH] drm/xe/vf: Retry GuC state reset until timeout

["Laguna, Lukasz" <lukasz.laguna@intel.com>]

On 8/31/2026 12:06, Michal Wajdeczko wrote:
>
> On 8/31/2026 11:21 AM, Lukasz Laguna wrote:
>> Instead of using a fixed number of GuC state reset retries, keep
>> retrying until timeout with sleeps between attempts.
> why?
> why 10s?

In case of VF FLR or PF FLR with VFs enabled, GuC reprovisioning may 
complete much later, so a fixed 10 retry window may not be enough.
Instead of relying on a fixed retry count, I wanted to introduce a 
deterministic maximum wait window, with sleeps between attempts, to 
avoid spamming requests while GuC is not ready.

Why 10s?

On the PF side, VF FLR uses a 5s timeout. I used 10s on the VF side to 
keep a safety margin from the VF perspective. This is empirical and may 
need adjustment, but so far I didn't observe any issues with this value.

I decided to send this now as a standalone reliability improvement 
(sleep + timeout), but maybe it'll be better justified when sent 
together with the FLR series, where the dependency and timing context 
are more explicit.

Lukasz

>
> you should describe the problem first
>
>> Signed-off-by: Lukasz Laguna <lukasz.laguna@intel.com>
>> ---
>>   drivers/gpu/drm/xe/xe_gt_sriov_vf.c | 12 +++++++++---
>>   1 file changed, 9 insertions(+), 3 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_vf.c b/drivers/gpu/drm/xe/xe_gt_sriov_vf.c
>> index 37899fcf5b22..01873896f77b 100644
>> --- a/drivers/gpu/drm/xe/xe_gt_sriov_vf.c
>> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_vf.c
>> @@ -31,6 +31,7 @@
>>   #include "xe_lrc.h"
>>   #include "xe_memirq.h"
>>   #include "xe_mmio.h"
>> +#include "xe_sleep.h"
>>   #include "xe_sriov.h"
>>   #include "xe_sriov_vf.h"
>>   #include "xe_sriov_vf_ccs.h"
>> @@ -86,18 +87,23 @@ static int guc_action_vf_reset(struct xe_guc *guc)
>>   	return ret > 0 ? -EPROTO : ret;
>>   }
>>   
>> -#define GUC_RESET_VF_STATE_RETRY_MAX	10
>> +#define GUC_RESET_VF_STATE_TIMEOUT_MS	10000
> 	10 * MSEC_PER_SEC /* 10s */
>
>> +#define GUC_RESET_VF_STATE_SLEEP_MS	10
>> +
>>   static int vf_reset_guc_state(struct xe_gt *gt)
>>   {
>> -	unsigned int retry = GUC_RESET_VF_STATE_RETRY_MAX;
>>   	struct xe_guc *guc = &gt->uc.guc;
>> +	unsigned long jiffies_end;
>>   	int err;
>>   
>> +	jiffies_end = jiffies + msecs_to_jiffies(GUC_RESET_VF_STATE_TIMEOUT_MS);
>>   	do {
>>   		err = guc_action_vf_reset(guc);
>>   		if (!err || err != -ETIMEDOUT)
>>   			break;
>> -	} while (--retry);
>> +
>> +		xe_sleep_relaxed_ms(GUC_RESET_VF_STATE_SLEEP_MS);
>> +	} while (time_before(jiffies, jiffies_end));
> btw, maybe we can use poll_timeout_us() here?
>
>>   
>>   	if (unlikely(err))
>>   		xe_gt_sriov_err(gt, "Failed to reset GuC state (%pe)\n", ERR_PTR(err));