[PATCH v2 26/32] drm/xe/madvise: Update migration policy based on preferred location

[Ghimiray, Himal Prasad]

On 15-05-2025 03:34, Matthew Brost wrote:
> On Mon, Apr 07, 2025 at 03:47:13PM +0530, Himal Prasad Ghimiray wrote:
>> When the user sets the valid devmem_fd as a preferred location, GPU fault
>> will trigger migration to tile of device associated with devmem_fd.
>>
>> If the user sets an invalid devmem_fd the preferred location is current
>> placement only.
>>
>> Signed-off-by: Himal Prasad Ghimiray <himal.prasad.ghimiray at intel.com>
>> ---
>>   drivers/gpu/drm/xe/xe_svm.c        | 15 ++++++++++++++-
>>   drivers/gpu/drm/xe/xe_vm.h         |  3 +++
>>   drivers/gpu/drm/xe/xe_vm_madvise.c | 20 +++++++++++++++++++-
>>   3 files changed, 36 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/xe/xe_svm.c b/drivers/gpu/drm/xe/xe_svm.c
>> index d40111e29bfe..60dfb1bf12ca 100644
>> --- a/drivers/gpu/drm/xe/xe_svm.c
>> +++ b/drivers/gpu/drm/xe/xe_svm.c
>> @@ -765,6 +765,12 @@ bool xe_svm_range_needs_migrate_to_vram(struct xe_svm_range *range, struct xe_vm
>>   	return needs_migrate;
>>   }
>>   
>> +static const u32 region_to_mem_type[] = {
>> +	XE_PL_TT,
>> +	XE_PL_VRAM0,
>> +	XE_PL_VRAM1,
>> +};
>> +
>>   /**
>>    * xe_svm_handle_pagefault() - SVM handle page fault
>>    * @vm: The VM.
>> @@ -796,6 +802,7 @@ int xe_svm_handle_pagefault(struct xe_vm *vm, struct xe_vma *vma,
>>   	struct xe_tile *tile = gt_to_tile(gt);
>>   	int retry_count = 3;
>>   	ktime_t end = 0;
>> +	u32 region;
>>   	int err;
>>   
>>   	lockdep_assert_held_write(&vm->lock);
>> @@ -820,7 +827,13 @@ int xe_svm_handle_pagefault(struct xe_vm *vm, struct xe_vma *vma,
>>   
>>   	range_debug(range, "PAGE FAULT");
>>   
>> -	if (xe_svm_range_needs_migrate_to_vram(range, vma, IS_DGFX(vm->xe))) {
>> +	region = vma->attr.preferred_loc.devmem_fd;
> 
> Mentioned this earlier in the series - you are assiging a devmem_fd to a
> region which is a bit confusing.

Hmm. Agreed

> 
>> +
>> +	if (xe_svm_range_needs_migrate_to_vram(range, vma, region)) {
>> +		region = region ? region : 1;
> 
> I think the default (region unset) should be the VRAM closest to the GT
> of the fault.

True

> 
>> +		/* Need rework for multigpu */
>> +		tile = &vm->xe->tiles[region_to_mem_type[region] - XE_PL_VRAM0];
>> +
>>   		err = xe_svm_alloc_vram(vm, tile, range, &ctx);
>>   		if (err) {
>>   			if (retry_count) {
>> diff --git a/drivers/gpu/drm/xe/xe_vm.h b/drivers/gpu/drm/xe/xe_vm.h
>> index 4e45230b7205..377f62f859b7 100644
>> --- a/drivers/gpu/drm/xe/xe_vm.h
>> +++ b/drivers/gpu/drm/xe/xe_vm.h
>> @@ -220,6 +220,9 @@ int __xe_vm_userptr_needs_repin(struct xe_vm *vm);
>>   
>>   int xe_vm_userptr_check_repin(struct xe_vm *vm);
>>   
>> +bool xe_vma_has_preferred_mem_loc(struct xe_vma *vma,
>> +				  u32 *mem_region, u32 *devmem_fd);
>> +
>>   int xe_vm_rebind(struct xe_vm *vm, bool rebind_worker);
>>   struct dma_fence *xe_vma_rebind(struct xe_vm *vm, struct xe_vma *vma,
>>   				u8 tile_mask);
>> diff --git a/drivers/gpu/drm/xe/xe_vm_madvise.c b/drivers/gpu/drm/xe/xe_vm_madvise.c
>> index 7e1a95106cb9..f870e8642190 100644
>> --- a/drivers/gpu/drm/xe/xe_vm_madvise.c
>> +++ b/drivers/gpu/drm/xe/xe_vm_madvise.c
>> @@ -61,7 +61,25 @@ static int madvise_preferred_mem_loc(struct xe_device *xe, struct xe_vm *vm,
>>   				     struct xe_vma **vmas, int num_vmas,
>>   				     struct drm_xe_madvise_ops ops)
>>   {
>> -	/* Implementation pending */
>> +	s32 devmem_fd;
>> +	u32 migration_policy;
>> +	int i;
>> +
>> +	xe_assert(vm->xe, ops.type == DRM_XE_VMA_ATTR_PREFERRED_LOC);
>> +	vm_dbg(&xe->drm, "migration policy = %d, devmem_fd = %d\n",
>> +	       ops.preferred_mem_loc.migration_policy,
>> +	       ops.preferred_mem_loc.devmem_fd);
> 
> As mentioned in patch #27, I'm not sure this debug info is all that
> useful.

Agreed. Planning to drop preferred location in next version and let it 
be brought in along with multi-device support.

> 
>> +
>> +	devmem_fd = (s32)ops.preferred_mem_loc.devmem_fd;
>> +	devmem_fd = (devmem_fd < 0) ? 0 : devmem_fd;
>> +
> 
> Why (devmem_fd < 0) ? 0? I'm not following this.

Incase of any -ve(invalid) fd being passed, wanted to have smem as 
preferred location.


> 
>> +	migration_policy = ops.preferred_mem_loc.migration_policy;
>> +
> 
> Mentioned earlier in the series, I'm confused by migration_policy and it
> also looks to be unused unless I'm missing something?

This was supposed to be used with multi-device support. >
> Matt
> 
>> +	for (i = 0; i < num_vmas; i++) {
>> +		vmas[i]->attr.preferred_loc.devmem_fd = devmem_fd;
>> +		vmas[i]->attr.preferred_loc.migration_policy = migration_policy;
>> +	}
>> +
>>   	return 0;
>>   }
>>   
>> -- 
>> 2.34.1
>>