Re: [PATCH v2 5/5] drm/pagemap: Enable THP support for GPU memory migration

[Matthew Brost <matthew.brost@intel.com>]

On Mon, Jan 05, 2026 at 12:18:28PM +0100, Francois Dugast wrote:
> This enables support for Transparent Huge Pages (THP) for device pages by
> using MIGRATE_VMA_SELECT_COMPOUND during migration. It removes the need to
> split folios and loop multiple times over all pages to perform required
> operations at page level. Instead, we rely on newly introduced support for
> higher orders in drm_pagemap and folio-level API.
> 
> In Xe, this drastically improves performance when using SVM. The GT stats
> below collected after a 2MB page fault show overall servicing is more than
> 7 times faster, and thanks to reduced CPU overhead the time spent on the
> actual copy goes from 23% without THP to 80% with THP:
> 
> Without THP:
> 
>     svm_2M_pagefault_us: 966
>     svm_2M_migrate_us: 942
>     svm_2M_device_copy_us: 223
>     svm_2M_get_pages_us: 9
>     svm_2M_bind_us: 10
> 
> With THP:
> 
>     svm_2M_pagefault_us: 132
>     svm_2M_migrate_us: 128
>     svm_2M_device_copy_us: 106
>     svm_2M_get_pages_us: 1
>     svm_2M_bind_us: 2
> 
> v2:
> - Fix one occurrence of drm_pagemap_get_devmem_page() (Matthew Brost)
> 
> Cc: Matthew Brost <matthew.brost@intel.com>
> Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
> Cc: Michal Mrozek <michal.mrozek@intel.com>
> Signed-off-by: Francois Dugast <francois.dugast@intel.com>
> ---
>  drivers/gpu/drm/drm_pagemap.c | 77 +++++++++++++++++++++++++++++++----
>  drivers/gpu/drm/xe/xe_svm.c   |  4 ++
>  include/drm/drm_pagemap.h     |  3 ++
>  3 files changed, 76 insertions(+), 8 deletions(-)
> 
> diff --git a/drivers/gpu/drm/drm_pagemap.c b/drivers/gpu/drm/drm_pagemap.c
> index 05e708730132..1ea8526ce946 100644
> --- a/drivers/gpu/drm/drm_pagemap.c
> +++ b/drivers/gpu/drm/drm_pagemap.c
> @@ -200,16 +200,20 @@ static void drm_pagemap_migration_unlock_put_pages(unsigned long npages,
>  /**
>   * drm_pagemap_get_devmem_page() - Get a reference to a device memory page
>   * @page: Pointer to the page
> + * @order: Order
>   * @zdd: Pointer to the GPU SVM zone device data
>   *
>   * This function associates the given page with the specified GPU SVM zone
>   * device data and initializes it for zone device usage.
>   */
>  static void drm_pagemap_get_devmem_page(struct page *page,
> +					unsigned int order,
>  					struct drm_pagemap_zdd *zdd)
>  {
> -	page->zone_device_data = drm_pagemap_zdd_get(zdd);
> -	zone_device_page_init(page, 0);
> +	struct folio *folio = page_folio(page);
> +
> +	folio_set_zone_device_data(folio, drm_pagemap_zdd_get(zdd));
> +	zone_device_page_init(page, order);
>  }
>  
>  /**
> @@ -522,7 +526,7 @@ int drm_pagemap_migrate_to_devmem(struct drm_pagemap_devmem *devmem_allocation,
>  		.end		= end,
>  		.pgmap_owner	= pagemap->owner,
>  		.flags		= MIGRATE_VMA_SELECT_SYSTEM | MIGRATE_VMA_SELECT_DEVICE_COHERENT |
> -		MIGRATE_VMA_SELECT_DEVICE_PRIVATE,
> +		MIGRATE_VMA_SELECT_DEVICE_PRIVATE | MIGRATE_VMA_SELECT_COMPOUND,
>  	};
>  	unsigned long i, npages = npages_in_range(start, end);
>  	unsigned long own_pages = 0, migrated_pages = 0;
> @@ -628,10 +632,13 @@ int drm_pagemap_migrate_to_devmem(struct drm_pagemap_devmem *devmem_allocation,
>  
>  	own_pages = 0;
>  
> -	for (i = 0; i < npages; ++i) {
> +	mutex_lock(&dpagemap->folio_split_lock);
> +	for (i = 0; i < npages;) {
> +		unsigned long j;
>  		struct page *page = pfn_to_page(migrate.dst[i]);
>  		struct page *src_page = migrate_pfn_to_page(migrate.src[i]);
>  		cur.start = i;
> +		unsigned int order;
>  
>  		pages[i] = NULL;
>  		if (src_page && is_device_private_page(src_page)) {
> @@ -658,7 +665,23 @@ int drm_pagemap_migrate_to_devmem(struct drm_pagemap_devmem *devmem_allocation,
>  			pages[i] = page;
>  		}
>  		migrate.dst[i] = migrate_pfn(migrate.dst[i]);
> -		drm_pagemap_get_devmem_page(page, zdd);
> +
> +		if (migrate.src[i] & MIGRATE_PFN_COMPOUND) {
> +			order = HPAGE_PMD_ORDER;

Can we assert the folio order is HPAGE_PMD_ORDER? Eventually mTHP could
be enabled on device folios where the order could be different than
HPAGE_PMD_ORDER. I suspect we will need some small updates the GPU SVM /
Xe to support mTHP and this assert would pop telling us we need to fix
GPU SVM.

> +
> +			migrate.dst[i] |= MIGRATE_PFN_COMPOUND;
> +
> +			for (j = 1; j < NR_PAGES(order) && i + j < npages; j++)
> +				migrate.dst[i + j] = 0;
> +
> +		} else {
> +			order = 0;
> +
> +			if (folio_order(page_folio(page)))
> +				migrate_device_split_page(page);
> +		}
> +
> +		drm_pagemap_get_devmem_page(page, order, zdd);
>  
>  		/* If we switched the migrating drm_pagemap, migrate previous pages now */
>  		err = drm_pagemap_migrate_range(devmem_allocation, migrate.src, migrate.dst,
> @@ -666,7 +689,11 @@ int drm_pagemap_migrate_to_devmem(struct drm_pagemap_devmem *devmem_allocation,
>  						mdetails);
>  		if (err)
>  			goto err_finalize;
> +
> +		i += NR_PAGES(order);
>  	}
> +	mutex_unlock(&dpagemap->folio_split_lock);
> +
>  	cur.start = npages;
>  	cur.ops = NULL; /* Force migration */
>  	err = drm_pagemap_migrate_range(devmem_allocation, migrate.src, migrate.dst,
> @@ -775,6 +802,8 @@ static int drm_pagemap_migrate_populate_ram_pfn(struct vm_area_struct *vas,
>  		page = folio_page(folio, 0);
>  		mpfn[i] = migrate_pfn(page_to_pfn(page));
>  
> +		if (order)
> +			mpfn[i] |= MIGRATE_PFN_COMPOUND;
>  next:
>  		if (page)
>  			addr += page_size(page);
> @@ -1030,8 +1059,15 @@ int drm_pagemap_evict_to_ram(struct drm_pagemap_devmem *devmem_allocation)
>  	if (err)
>  		goto err_finalize;
>  
> -	for (i = 0; i < npages; ++i)
> +	for (i = 0; i < npages;) {
> +		unsigned int order = 0;
> +
>  		pages[i] = migrate_pfn_to_page(src[i]);
> +		if (pages[i])
> +			order = folio_order(page_folio(pages[i]));
> +
> +		i += NR_PAGES(order);
> +	}
>  
>  	err = ops->copy_to_ram(pages, pagemap_addr, npages, NULL);
>  	if (err)
> @@ -1084,7 +1120,8 @@ static int __drm_pagemap_migrate_to_ram(struct vm_area_struct *vas,
>  		.vma		= vas,
>  		.pgmap_owner	= page_pgmap(page)->owner,
>  		.flags		= MIGRATE_VMA_SELECT_DEVICE_PRIVATE |
> -		MIGRATE_VMA_SELECT_DEVICE_COHERENT,
> +		MIGRATE_VMA_SELECT_DEVICE_COHERENT |
> +		MIGRATE_VMA_SELECT_COMPOUND,

The alignment was off here - e.g. my vi settings make the original code
look like this:

                .flags          = MIGRATE_VMA_SELECT_DEVICE_PRIVATE |
                        MIGRATE_VMA_SELECT_DEVICE_COHERENT,

Since you are changing this line can you fix the alignment?

>  		.fault_page	= page,
>  	};
>  	struct drm_pagemap_migrate_details mdetails = {};
> @@ -1150,8 +1187,15 @@ static int __drm_pagemap_migrate_to_ram(struct vm_area_struct *vas,
>  	if (err)
>  		goto err_finalize;
>  
> -	for (i = 0; i < npages; ++i)
> +	for (i = 0; i < npages;) {
> +		unsigned int order = 0;
> +
>  		pages[i] = migrate_pfn_to_page(migrate.src[i]);
> +		if (pages[i])
> +			order = folio_order(page_folio(pages[i]));
> +
> +		i += NR_PAGES(order);
> +	}
>  
>  	err = ops->copy_to_ram(pages, pagemap_addr, npages, NULL);
>  	if (err)
> @@ -1209,9 +1253,26 @@ static vm_fault_t drm_pagemap_migrate_to_ram(struct vm_fault *vmf)
>  	return err ? VM_FAULT_SIGBUS : 0;
>  }
>  
> +static void drm_pagemap_folio_split(struct folio *orig_folio, struct folio *new_folio)
> +{
> +	struct drm_pagemap_zdd *zdd;
> +
> +	if (!new_folio)
> +		return;
> +
> +	new_folio->pgmap = orig_folio->pgmap;
> +	zdd = folio_zone_device_data(orig_folio);
> +	if (folio_order(new_folio))
> +		folio_set_zone_device_data(new_folio, drm_pagemap_zdd_get(zdd));
> +	else
> +		folio_page(new_folio, 0)->zone_device_data =
> +			drm_pagemap_zdd_get(zdd);

Is this if statement needed? I believe folio_set_zone_device_data can
just be called.

> +}
> +
>  static const struct dev_pagemap_ops drm_pagemap_pagemap_ops = {
>  	.folio_free = drm_pagemap_folio_free,
>  	.migrate_to_ram = drm_pagemap_migrate_to_ram,
> +	.folio_split = drm_pagemap_folio_split,
>  };
>  
>  /**
> diff --git a/drivers/gpu/drm/xe/xe_svm.c b/drivers/gpu/drm/xe/xe_svm.c
> index fa2ee2c08f31..05dba6abbcc8 100644
> --- a/drivers/gpu/drm/xe/xe_svm.c
> +++ b/drivers/gpu/drm/xe/xe_svm.c
> @@ -1760,6 +1760,10 @@ static struct xe_pagemap *xe_pagemap_create(struct xe_device *xe, struct xe_vram
>  	if (err)
>  		goto out_no_dpagemap;
>  
> +	err = drmm_mutex_init(&xe->drm, &dpagemap->folio_split_lock);
> +	if (err)
> +		goto out_err;
> +
>  	res = devm_request_free_mem_region(dev, &iomem_resource,
>  					   vr->usable_size);
>  	if (IS_ERR(res)) {
> diff --git a/include/drm/drm_pagemap.h b/include/drm/drm_pagemap.h
> index 736fb6cb7b33..3c8bacfc79e6 100644
> --- a/include/drm/drm_pagemap.h
> +++ b/include/drm/drm_pagemap.h
> @@ -161,6 +161,7 @@ struct drm_pagemap_ops {
>   * &struct drm_pagemap. May be NULL if no cache is used.
>   * @shrink_link: Link into the shrinker's list of drm_pagemaps. Only
>   * used if also using a pagemap cache.
> + * @folio_split_lock: Lock to protect device folio splitting.
>   */
>  struct drm_pagemap {
>  	const struct drm_pagemap_ops *ops;
> @@ -170,6 +171,8 @@ struct drm_pagemap {
>  	struct drm_pagemap_dev_hold *dev_hold;
>  	struct drm_pagemap_cache *cache;
>  	struct list_head shrink_link;
> +	/* Protect device folio splitting */

I don't think you need this comment as we have kernel doc.

Nits aside, patch looks good.

Matt

> +	struct mutex folio_split_lock;
>  };
>  
>  struct drm_pagemap_devmem;
> -- 
> 2.43.0
>