Re: [RFC PATCH] drm/xe/lnl: Implement clear-on-free for pooled BOs

["Thomas Hellström" <thomas.hellstrom@linux.intel.com>]

Hi, Nirmoy,

On Thu, 2024-08-22 at 14:42 +0200, Nirmoy Das wrote:
> Implement GPU clear-on-free for pooled system pages in Xe.
> 
> Ensure proper use of TTM_TT_FLAG_CLEARED_ON_FREE by leveraging
> ttm_device_funcs.release_notify() for GPU clear-on-free. If GPU clear
> fails, xe_ttm_tt_unpopulate() will fallback to CPU clear.
> 
> Clear-on-free is only relevant for pooled pages as driver needs to
> give
> back those pages. So do clear-on-free only for such BOs and keep
> doing
> clear-on-alloc for ttm_cached type BOs
> 
> Cc: Matthew Auld <matthew.auld@intel.com>
> Cc: Matthew Brost <matthew.brost@intel.com>
> Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
> Signed-off-by: Nirmoy Das <nirmoy.das@intel.com>

While this would probably work, I don't immediately see the benefit
over CPU clearing, since we have no way of combining this with the CCS
clear, right? So the clearing latency will most probably be increased,
but the bo releasing thread won't see that because the waiting for
clear is offloaded to the TTM delayed destroy mechanism.

Also, once we've dropped the gem refcount to zero, the gem members of
the object, including bo_move, are strictly not valid anymore and
shouldn't be used.

If we want to try to improve freeing latency by offloading the clearing
on free to a separate CPU thread, though, maybe we could discuss with
Christian to always (or if a flag in the ttm device requests it) take
the TTM delayed destruction path for bos with pooled pages, rather than
to free them sync, something along the lines of:

diff --git a/drivers/gpu/drm/ttm/ttm_bo.c
b/drivers/gpu/drm/ttm/ttm_bo.c
index 320592435252..fca69ec1740d 100644
--- a/drivers/gpu/drm/ttm/ttm_bo.c
+++ b/drivers/gpu/drm/ttm/ttm_bo.c
@@ -271,7 +271,7 @@ static void ttm_bo_release(struct kref *kref)
 
                if (!dma_resv_test_signaled(bo->base.resv,
                                            DMA_RESV_USAGE_BOOKKEEP) ||
-                   (want_init_on_free() && (bo->ttm != NULL)) ||
+                   (bo->ttm && (want_init_on_free() || bo->ttm-
>caching != ttm_cached)) ||
                    bo->type == ttm_bo_type_sg ||
                    !dma_resv_trylock(bo->base.resv)) {
                        /* The BO is not idle, resurrect it for delayed
destroy */

Would ofc require some substantial proven latency gain, though. Overall
system cpu usage would probably not improve. 

/Thomas


> ---
>  drivers/gpu/drm/xe/xe_bo.c | 101 +++++++++++++++++++++++++++++++++--
> --
>  1 file changed, 91 insertions(+), 10 deletions(-)
> 
> diff --git a/drivers/gpu/drm/xe/xe_bo.c b/drivers/gpu/drm/xe/xe_bo.c
> index 6ed0e1955215..e7bc74f8ae82 100644
> --- a/drivers/gpu/drm/xe/xe_bo.c
> +++ b/drivers/gpu/drm/xe/xe_bo.c
> @@ -283,6 +283,8 @@ struct xe_ttm_tt {
>  	struct device *dev;
>  	struct sg_table sgt;
>  	struct sg_table *sg;
> +	bool sys_clear_on_free;
> +	bool sys_clear_on_alloc;
>  };
>  
>  static int xe_tt_map_sg(struct ttm_tt *tt)
> @@ -401,8 +403,23 @@ static struct ttm_tt *xe_ttm_tt_create(struct
> ttm_buffer_object *ttm_bo,
>  	 * flag. Zeroed pages are only required for
> ttm_bo_type_device so
>  	 * unwanted data is not leaked to userspace.
>  	 */
> -	if (ttm_bo->type == ttm_bo_type_device && xe-
> >mem.gpu_page_clear_sys)
> -		page_flags |= TTM_TT_FLAG_CLEARED_ON_FREE;
> +	if (ttm_bo->type == ttm_bo_type_device && xe-
> >mem.gpu_page_clear_sys) {
> +		/*
> +		 * Non-pooled BOs are always clear on alloc when
> possible.
> +		 * clear-on-free is not needed as there is no pool
> to give pages back.
> +		 */
> +		if (caching == ttm_cached) {
> +			tt->sys_clear_on_alloc = true;
> +			tt->sys_clear_on_free = false;
> +		} else {
> +		/*
> +		 * For pooled BO, clear-on-alloc is done by the CPU
> for now and
> +		 * GPU will do clear on free when releasing the BO.
> +		 */
> +			tt->sys_clear_on_alloc = false;
> +			tt->sys_clear_on_free = true;
> +		}
> +	}
>  
>  	err = ttm_tt_init(&tt->ttm, &bo->ttm, page_flags, caching,
> extra_pages);
>  	if (err) {
> @@ -416,8 +433,10 @@ static struct ttm_tt *xe_ttm_tt_create(struct
> ttm_buffer_object *ttm_bo,
>  static int xe_ttm_tt_populate(struct ttm_device *ttm_dev, struct
> ttm_tt *tt,
>  			      struct ttm_operation_ctx *ctx)
>  {
> +	struct xe_ttm_tt *xe_tt;
>  	int err;
>  
> +	xe_tt = container_of(tt, struct xe_ttm_tt, ttm);
>  	/*
>  	 * dma-bufs are not populated with pages, and the dma-
>  	 * addresses are set up when moved to XE_PL_TT.
> @@ -426,7 +445,7 @@ static int xe_ttm_tt_populate(struct ttm_device
> *ttm_dev, struct ttm_tt *tt,
>  		return 0;
>  
>  	/* Clear TTM_TT_FLAG_ZERO_ALLOC when GPU is set to clear
> system pages */
> -	if (tt->page_flags & TTM_TT_FLAG_CLEARED_ON_FREE)
> +	if (xe_tt->sys_clear_on_alloc)
>  		tt->page_flags &= ~TTM_TT_FLAG_ZERO_ALLOC;
>  
>  	err = ttm_pool_alloc(&ttm_dev->pool, tt, ctx);
> @@ -438,11 +457,19 @@ static int xe_ttm_tt_populate(struct ttm_device
> *ttm_dev, struct ttm_tt *tt,
>  
>  static void xe_ttm_tt_unpopulate(struct ttm_device *ttm_dev, struct
> ttm_tt *tt)
>  {
> +	struct xe_ttm_tt *xe_tt;
> +
> +	xe_tt = container_of(tt, struct xe_ttm_tt, ttm);
> +
>  	if (tt->page_flags & TTM_TT_FLAG_EXTERNAL)
>  		return;
>  
>  	xe_tt_unmap_sg(tt);
>  
> +	/* Hint TTM pool that pages are already cleared */
> +	if (xe_tt->sys_clear_on_free)
> +		tt->page_flags |= TTM_TT_FLAG_CLEARED_ON_FREE;
> +
>  	return ttm_pool_free(&ttm_dev->pool, tt);
>  }
>  
> @@ -664,6 +691,7 @@ static int xe_bo_move(struct ttm_buffer_object
> *ttm_bo, bool evict,
>  	struct ttm_resource *old_mem = ttm_bo->resource;
>  	u32 old_mem_type = old_mem ? old_mem->mem_type :
> XE_PL_SYSTEM;
>  	struct ttm_tt *ttm = ttm_bo->ttm;
> +	struct xe_ttm_tt *xe_tt;
>  	struct xe_migrate *migrate = NULL;
>  	struct dma_fence *fence;
>  	bool move_lacks_source;
> @@ -674,12 +702,13 @@ static int xe_bo_move(struct ttm_buffer_object
> *ttm_bo, bool evict,
>  	bool clear_system_pages;
>  	int ret = 0;
>  
> +	xe_tt = container_of(ttm_bo->ttm, struct xe_ttm_tt, ttm);
>  	/*
>  	 * Clear TTM_TT_FLAG_CLEARED_ON_FREE on bo creation path
> when
>  	 * moving to system as the bo doesn't have dma_mapping.
>  	 */
>  	if (!old_mem && ttm && !ttm_tt_is_populated(ttm))
> -		ttm->page_flags &= ~TTM_TT_FLAG_CLEARED_ON_FREE;
> +		xe_tt->sys_clear_on_alloc = false;
>  
>  	/* Bo creation path, moving to system or TT. */
>  	if ((!old_mem && ttm) && !handle_system_ccs) {
> @@ -703,10 +732,9 @@ static int xe_bo_move(struct ttm_buffer_object
> *ttm_bo, bool evict,
>  	move_lacks_source = handle_system_ccs ? (!bo->ccs_cleared) 
> :
>  						(!mem_type_is_vram(o
> ld_mem_type) && !tt_has_data);
>  
> -	clear_system_pages = ttm && (ttm->page_flags &
> TTM_TT_FLAG_CLEARED_ON_FREE);
> +	clear_system_pages = ttm && xe_tt->sys_clear_on_alloc;
>  	needs_clear = (ttm && ttm->page_flags &
> TTM_TT_FLAG_ZERO_ALLOC) ||
> -		(!ttm && ttm_bo->type == ttm_bo_type_device) ||
> -		clear_system_pages;
> +		(!ttm && ttm_bo->type == ttm_bo_type_device) ||
> clear_system_pages;
>  
>  	if (new_mem->mem_type == XE_PL_TT) {
>  		ret = xe_tt_map_sg(ttm);
> @@ -1028,10 +1056,47 @@ static bool
> xe_ttm_bo_lock_in_destructor(struct ttm_buffer_object *ttm_bo)
>  	return locked;
>  }
>  
> +static struct dma_fence *xe_ttm_bo_clear_on_free(struct
> ttm_buffer_object *ttm_bo)
> +{
> +	struct xe_bo *bo  = ttm_to_xe_bo(ttm_bo);
> +	struct xe_device *xe = xe_bo_device(bo);
> +	struct xe_migrate *migrate;
> +	struct xe_ttm_tt *xe_tt;
> +	struct dma_fence *clear_fence;
> +
> +	/* return early if nothing to clear */
> +	if (!ttm_bo->ttm)
> +		return NULL;
> +
> +	xe_tt = container_of(ttm_bo->ttm, struct xe_ttm_tt, ttm);
> +	/* return early if nothing to clear */
> +	if (!xe_tt->sys_clear_on_free || !bo->ttm.resource)
> +		return NULL;
> +
> +	if (XE_WARN_ON(!xe_tt->sg))
> +		return NULL;
> +
> +	if (bo->tile)
> +		migrate = bo->tile->migrate;
> +	else
> +		migrate = xe->tiles[0].migrate;
> +
> +	xe_assert(xe, migrate);
> +
> +	clear_fence = xe_migrate_clear(migrate, bo, bo-
> >ttm.resource,
> +				       XE_MIGRATE_CLEAR_FLAG_FULL);
> +	if (IS_ERR(clear_fence))
> +		return NULL;
> +
> +	xe_tt->sys_clear_on_free = false;
> +
> +	return clear_fence;
> +}
> +
>  static void xe_ttm_bo_release_notify(struct ttm_buffer_object
> *ttm_bo)
>  {
>  	struct dma_resv_iter cursor;
> -	struct dma_fence *fence;
> +	struct dma_fence *clear_fence, *fence;
>  	struct dma_fence *replacement = NULL;
>  	struct xe_bo *bo;
>  
> @@ -1041,15 +1106,31 @@ static void xe_ttm_bo_release_notify(struct
> ttm_buffer_object *ttm_bo)
>  	bo = ttm_to_xe_bo(ttm_bo);
>  	xe_assert(xe_bo_device(bo), !(bo->created &&
> kref_read(&ttm_bo->base.refcount)));
>  
> +	clear_fence = xe_ttm_bo_clear_on_free(ttm_bo);
> +
>  	/*
>  	 * Corner case where TTM fails to allocate memory and this
> BOs resv
>  	 * still points the VMs resv
>  	 */
> -	if (ttm_bo->base.resv != &ttm_bo->base._resv)
> +	if (ttm_bo->base.resv != &ttm_bo->base._resv) {
> +		if (clear_fence)
> +			dma_fence_wait(clear_fence, false);
>  		return;
> +	}
>  
> -	if (!xe_ttm_bo_lock_in_destructor(ttm_bo))
> +	if (!xe_ttm_bo_lock_in_destructor(ttm_bo)) {
> +		if (clear_fence)
> +			dma_fence_wait(clear_fence, false);
>  		return;
> +	}
> +
> +	if (clear_fence) {
> +		if (dma_resv_reserve_fences(ttm_bo->base.resv, 1))
> +			dma_fence_wait(clear_fence, false);
> +		else
> +			dma_resv_add_fence(ttm_bo->base.resv,
> clear_fence,
> +					   DMA_RESV_USAGE_KERNEL);
> +	}
>  
>  	/*
>  	 * Scrub the preempt fences if any. The unbind fence is
> already