[Intel-gfx] [PATCH 19/19] drm/i915: Add accelerated migration to ttm
Thomas Hellström (Intel)
thomas_os at shipmail.org
Thu Sep 16 11:19:26 UTC 2021
drm/i915/ttm: Add async migration to the TTM backend?
On 8/30/21 2:10 PM, Maarten Lankhorst wrote:
> Expose the fence to ttm_bo->moving, which will get picked up by i915
> through the i915_gem_object_get_moving_fence call. Should be sufficient
> for the needs we have.
>
> Signed-off-by: Maarten Lankhorst <maarten.lankhorst at linux.intel.com>
> ---
> drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 104 ++++++++++++------------
> 1 file changed, 54 insertions(+), 50 deletions(-)
>
> diff --git a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c
> index eaf2ff29dd4a..751abbc3ccd3 100644
> --- a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c
> +++ b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c
> @@ -353,10 +353,14 @@ static void i915_ttm_swap_notify(struct ttm_buffer_object *bo)
>
> static void i915_ttm_delete_mem_notify(struct ttm_buffer_object *bo)
> {
> - struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
> + struct drm_i915_gem_object *obj;
>
> - if (likely(obj))
> - i915_ttm_free_cached_io_st(obj);
> + if (bo->destroy != i915_ttm_bo_destroy)
> + return;
> +
> + obj = i915_ttm_to_gem(bo);
> +
> + i915_ttm_free_cached_io_st(obj);
> }
>
> static struct intel_memory_region *
> @@ -426,10 +430,10 @@ i915_ttm_resource_get_st(struct drm_i915_gem_object *obj,
> return intel_region_ttm_resource_to_st(obj->mm.region, res);
> }
>
> -static int i915_ttm_accel_move(struct ttm_buffer_object *bo,
> - bool clear,
> - struct ttm_resource *dst_mem,
> - struct sg_table *dst_st)
> +static struct i915_request *i915_ttm_accel_move(struct ttm_buffer_object *bo,
> + bool clear,
> + struct ttm_resource *dst_mem,
> + struct sg_table *dst_st)
> {
> struct drm_i915_private *i915 = container_of(bo->bdev, typeof(*i915),
> bdev);
> @@ -437,18 +441,18 @@ static int i915_ttm_accel_move(struct ttm_buffer_object *bo,
> ttm_manager_type(bo->bdev, bo->resource->mem_type);
> struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
> struct sg_table *src_st;
> - struct i915_request *rq;
> + struct i915_request *rq = NULL;
> struct ttm_tt *ttm = bo->ttm;
> enum i915_cache_level src_level, dst_level;
> int ret;
>
> if (!i915->gt.migrate.context)
> - return -EINVAL;
> + return ERR_PTR(-EINVAL);
>
> dst_level = i915_ttm_cache_level(i915, dst_mem, ttm);
> if (clear) {
> if (bo->type == ttm_bo_type_kernel)
> - return -EINVAL;
> + return ERR_PTR(-EINVAL);
>
> intel_engine_pm_get(i915->gt.migrate.context->engine);
> ret = intel_context_migrate_clear(i915->gt.migrate.context, NULL,
In the eviction or migration case, do we rely on unbind to sync on
previous rendering?
> @@ -456,10 +460,6 @@ static int i915_ttm_accel_move(struct ttm_buffer_object *bo,
> gpu_binds_iomem(dst_mem),
> 0, &rq);
>
> - if (!ret && rq) {
> - i915_request_wait(rq, 0, MAX_SCHEDULE_TIMEOUT);
> - i915_request_put(rq);
> - }
> intel_engine_pm_put(i915->gt.migrate.context->engine);
> } else {
> src_st = src_man->use_tt ? i915_ttm_tt_get_st(ttm) :
> @@ -473,49 +473,40 @@ static int i915_ttm_accel_move(struct ttm_buffer_object *bo,
> dst_st->sgl, dst_level,
> gpu_binds_iomem(dst_mem),
> &rq);
> - if (!ret && rq) {
> - i915_request_wait(rq, 0, MAX_SCHEDULE_TIMEOUT);
> - i915_request_put(rq);
> - }
> intel_engine_pm_put(i915->gt.migrate.context->engine);
> }
>
> - return ret;
> + return ret ? ERR_PTR(ret) : rq;
Actually, IIRC, rq can be Non-NULL even if we hit an error, in which
case the request blit payload is ditched. Not sure what the best action
is in that case, but perhaps just sync and unref. Perhaps we don't even
need to sync.
> }
>
> -static void __i915_ttm_move(struct ttm_buffer_object *bo, bool clear,
> +static void i915_ttm_memcpy(struct ttm_buffer_object *bo, bool clear,
> struct ttm_resource *dst_mem,
> struct sg_table *dst_st)
> {
> - int ret;
> + struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
> + struct intel_memory_region *dst_reg, *src_reg;
> + union {
> + struct ttm_kmap_iter_tt tt;
> + struct ttm_kmap_iter_iomap io;
> + } _dst_iter, _src_iter;
> + struct ttm_kmap_iter *dst_iter, *src_iter;
>
> - ret = i915_ttm_accel_move(bo, clear, dst_mem, dst_st);
> - if (ret) {
> - struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
> - struct intel_memory_region *dst_reg, *src_reg;
> - union {
> - struct ttm_kmap_iter_tt tt;
> - struct ttm_kmap_iter_iomap io;
> - } _dst_iter, _src_iter;
> - struct ttm_kmap_iter *dst_iter, *src_iter;
> -
> - dst_reg = i915_ttm_region(bo->bdev, dst_mem->mem_type);
> - src_reg = i915_ttm_region(bo->bdev, bo->resource->mem_type);
> - GEM_BUG_ON(!dst_reg || !src_reg);
> -
> - dst_iter = !cpu_maps_iomem(dst_mem) ?
> - ttm_kmap_iter_tt_init(&_dst_iter.tt, bo->ttm) :
> - ttm_kmap_iter_iomap_init(&_dst_iter.io, &dst_reg->iomap,
> - dst_st, dst_reg->region.start);
> -
> - src_iter = !cpu_maps_iomem(bo->resource) ?
> - ttm_kmap_iter_tt_init(&_src_iter.tt, bo->ttm) :
> - ttm_kmap_iter_iomap_init(&_src_iter.io, &src_reg->iomap,
> - obj->ttm.cached_io_st,
> - src_reg->region.start);
> -
> - ttm_move_memcpy(clear, dst_mem->num_pages, dst_iter, src_iter);
> - }
> + dst_reg = i915_ttm_region(bo->bdev, dst_mem->mem_type);
> + src_reg = i915_ttm_region(bo->bdev, bo->resource->mem_type);
> + GEM_BUG_ON(!dst_reg || !src_reg);
> +
> + dst_iter = !cpu_maps_iomem(dst_mem) ?
> + ttm_kmap_iter_tt_init(&_dst_iter.tt, bo->ttm) :
> + ttm_kmap_iter_iomap_init(&_dst_iter.io, &dst_reg->iomap,
> + dst_st, dst_reg->region.start);
> +
> + src_iter = !cpu_maps_iomem(bo->resource) ?
> + ttm_kmap_iter_tt_init(&_src_iter.tt, bo->ttm) :
> + ttm_kmap_iter_iomap_init(&_src_iter.io, &src_reg->iomap,
> + obj->ttm.cached_io_st,
> + src_reg->region.start);
> +
> + ttm_move_memcpy(clear, dst_mem->num_pages, dst_iter, src_iter);
> }
>
> static int i915_ttm_move(struct ttm_buffer_object *bo, bool evict,
> @@ -530,6 +521,7 @@ static int i915_ttm_move(struct ttm_buffer_object *bo, bool evict,
> struct sg_table *dst_st;
> bool clear;
> int ret;
> + struct i915_request *rq;
>
> /* Sync for now. We could do the actual copy async. */
> ret = ttm_bo_wait_ctx(bo, ctx);
> @@ -558,10 +550,22 @@ static int i915_ttm_move(struct ttm_buffer_object *bo, bool evict,
> return PTR_ERR(dst_st);
>
> clear = !cpu_maps_iomem(bo->resource) && (!ttm || !ttm_tt_is_populated(ttm));
> +
> + rq = NULL;
> if (!(clear && ttm && !(ttm->page_flags & TTM_PAGE_FLAG_ZERO_ALLOC)))
> - __i915_ttm_move(bo, clear, dst_mem, dst_st);
> + rq = i915_ttm_accel_move(bo, clear, dst_mem, dst_st);
> +
> + if (IS_ERR(rq))
> + i915_ttm_memcpy(bo, clear, dst_mem, dst_st);
> +
> + /* Below dst_mem becomes bo->resource. */
> + if (!IS_ERR_OR_NULL(rq)) {
> + ttm_bo_move_accel_cleanup(bo, &rq->fence, evict, true, dst_mem);
> + i915_request_put(rq);
> + } else {
> + ttm_bo_move_sync_cleanup(bo, dst_mem);
> + }
>
> - ttm_bo_move_sync_cleanup(bo, dst_mem);
> i915_ttm_adjust_domains_after_move(obj);
> i915_ttm_free_cached_io_st(obj);
>
More information about the Intel-gfx
mailing list