Re: [PATCH v3 3/5] drm/xe/vf: Skip fixups on VF migration before getting GGTT info

["Lis, Tomasz" <tomasz.lis@intel.com>]

On 10/17/2025 12:20 AM, Michal Wajdeczko wrote:
>
> On 10/16/2025 2:05 PM, Tomasz Lis wrote:
>> The GuC RESFIX state should be achievable only after a successful
>> handshake. If VF KMD has no GGTT configuration yet and we still got
>> into RESFIX state, then either we're dealing with unclean initial
>> state due to unusual actions before probe, or the migration
>> happened while xe init (started by probe) was running.
>>
>> In 1st case (VF migration before probe), we should just skip migration.
>> Init procedure will ensure exit from RESFIX state as it starts GuC
>> handshake with a reset.
>>
>> In 2nd case (VF migration during xe init), the migration procedure
>> should execute normally if GGTT configuration was already acquired
>> from GuC, and can be skipped if it was not acquired.
> we initiate recovery as part of the MEMIRQ handling
>
> but to get MEMIRQ interrupt from the GuC we need to setup it first
> to setup MEMIRQ vector we need to obtain GGTT data from GuC
> only the we can register this MEMIRQ vector in GuC
>
> so valid GGTT data seems to be hard prerequisite for this scenario
>
> thus vf_ggtt_queried() should be always true, no?

True. Patch makes little sense. Will discard.

-Tomasz

>
>> This solution will avoid crashes due to the VF migration running
>> on non-initialized xe sub-structures. But it is not enough to allow
>> fully reliable migration during driver probe. In particular, the
>> situation where the probe might not end successfully, is:
>>
>> * The VF is paused and migrated after GuC reset (vf_bootstrap) but
>> before config is acquired (vf_query_config). In such case, GuC may
>> remain in RESFIX state, leading to timeouting requests.
>>
>> Signed-off-by: Tomasz Lis <tomasz.lis@intel.com>
>> ---
>>   drivers/gpu/drm/xe/xe_gt_sriov_vf.c | 12 ++++++++++++
>>   1 file changed, 12 insertions(+)
>>
>> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_vf.c b/drivers/gpu/drm/xe/xe_gt_sriov_vf.c
>> index 34c68de6e2f3..bb0b71a47125 100644
>> --- a/drivers/gpu/drm/xe/xe_gt_sriov_vf.c
>> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_vf.c
>> @@ -1149,6 +1149,12 @@ void xe_gt_sriov_vf_print_version(struct xe_gt *gt, struct drm_printer *p)
>>   		   pf_version->major, pf_version->minor);
>>   }
>>   
>> +static bool vf_ggtt_queried(struct xe_tile *tile)
>> +{
>> +	guard(mutex)(&tile->mem.ggtt->lock);
>> +	return xe_tile_sriov_vf_ggtt(tile) != 0;
>> +}
>> +
>>   static bool vf_post_migration_shutdown(struct xe_gt *gt)
>>   {
>>   	struct xe_device *xe = gt_to_xe(gt);
>> @@ -1260,6 +1266,11 @@ static void vf_post_migration_recovery(struct xe_gt *gt)
>>   	xe_gt_sriov_dbg(gt, "migration recovery in progress\n");
>>   
>>   	xe_pm_runtime_get(xe);
>> +
>> +	/* If during init and before GGTT configuration, skip the procedure. */
>> +	if (!vf_ggtt_queried(gt_to_tile(gt)))
>> +		goto skip;
>> +
>>   	retry = vf_post_migration_shutdown(gt);
>>   	if (retry)
>>   		goto queue;
>> @@ -1282,6 +1293,7 @@ static void vf_post_migration_recovery(struct xe_gt *gt)
>>   
>>   	vf_post_migration_kickstart(gt);
>>   
>> +skip:
>>   	xe_pm_runtime_put(xe);
>>   	xe_gt_sriov_notice(gt, "migration recovery ended\n");
>>   	return;