[PATCH v3 2/2] drm/xe: Wait for migration job before unmapping pages

[Nirmoy Das]

On 12/10/2024 5:39 PM, Thomas Hellström wrote:
> On Tue, 2024-12-10 at 17:15 +0100, Nirmoy Das wrote:
>> Fix a potential GPU page fault during tt -> system moves by waiting
>> for
>> migration jobs to complete before unmapping SG. This ensures that
>> IOMMU
>> mappings are not prematurely torn down while a migration job is still
>> in
>> progress.
>>
>> v2: Use intr=false(Matt A)
>> v3: Update commit message(Matt A)
>>
>> Closes: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/3466
>> Fixes: 75521e8b56e8 ("drm/xe: Perform dma_map when moving system
>> buffer objects to TT")
>> Cc: Thomas Hellström <thomas.hellstrom at linux.intel.com>
>> Cc: Matthew Brost <matthew.brost at intel.com>
>> Cc: Lucas De Marchi <lucas.demarchi at intel.com>
>> Cc: <stable at vger.kernel.org> # v6.11+
>> Cc: Matthew Auld <matthew.auld at intel.com>
>> Signed-off-by: Nirmoy Das <nirmoy.das at intel.com>
>> Reviewed-by: Matthew Auld <matthew.auld at intel.com>
>> ---
>>  drivers/gpu/drm/xe/xe_bo.c | 10 +++++++++-
>>  1 file changed, 9 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/gpu/drm/xe/xe_bo.c b/drivers/gpu/drm/xe/xe_bo.c
>> index 06931df876ab..0a41b6c0583a 100644
>> --- a/drivers/gpu/drm/xe/xe_bo.c
>> +++ b/drivers/gpu/drm/xe/xe_bo.c
>> @@ -857,8 +857,16 @@ static int xe_bo_move(struct ttm_buffer_object
>> *ttm_bo, bool evict,
>>  
>>  out:
>>  	if ((!ttm_bo->resource || ttm_bo->resource->mem_type ==
>> XE_PL_SYSTEM) &&
>> -	    ttm_bo->ttm)
>> +	    ttm_bo->ttm) {
>> +		long timeout = dma_resv_wait_timeout(ttm_bo-
>>> base.resv,
>> +						    
>> DMA_RESV_USAGE_BOOKKEEP,
>> +						     false,
>> +						    
>> MAX_SCHEDULE_TIMEOUT);
>> +		if (timeout < 0)
>> +			ret = timeout;
>> +
>>  		xe_tt_unmap_sg(ttm_bo->ttm);
>> +	}
>>  
>>  	return ret;
>>  }
> I assume here we're waiting for the move fence, right? 
Yes
> However if
> @evict is true, we should hit the ttm_bo_wait_free_node() path. In what
> cases do we hit this without evict being true?
evict is false here. It is coming from xe_bo_evict() rather than ttm_mem_evict_*

Call flow looks like: ccs_test_migrate()->xe_bo_evict()->ttm_bo_validate()->ttm_bo_handle_move_mem()->xe_bo_move()



>
> Also, shouldn't it be sufficient to wait for DMA_RESV_USAGE_KERNEL
> here? 

Yes, DMA_RESV_USAGE_KERNEL should be enough as we are waiting for the migration fence.


Regards,

Nirmoy

>
> Thanks,
> Thomas
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/intel-xe/attachments/20241210/dab473ba/attachment.htm>