[Intel-gfx] [PATCH 5/5] drm/i915: Tidy batch pool logic
John Harrison
John.C.Harrison at Intel.com
Fri Feb 13 06:00:50 PST 2015
On 14/01/2015 11:20, Chris Wilson wrote:
> Move the madvise logic out of the execbuffer main path into the
> relatively rare allocation path, making the execbuffer manipulation less
> fragile.
>
> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> ---
> drivers/gpu/drm/i915/i915_cmd_parser.c | 12 +++---------
> drivers/gpu/drm/i915/i915_gem_batch_pool.c | 31 +++++++++++++++---------------
> drivers/gpu/drm/i915/i915_gem_execbuffer.c | 11 +++--------
> 3 files changed, 21 insertions(+), 33 deletions(-)
>
> diff --git a/drivers/gpu/drm/i915/i915_cmd_parser.c b/drivers/gpu/drm/i915/i915_cmd_parser.c
> index 9a6da3536ae5..3e5e8cb54a88 100644
> --- a/drivers/gpu/drm/i915/i915_cmd_parser.c
> +++ b/drivers/gpu/drm/i915/i915_cmd_parser.c
> @@ -866,6 +866,9 @@ static u32 *copy_batch(struct drm_i915_gem_object *dest_obj,
> batch_len + batch_start_offset > src_obj->base.size)
> return ERR_PTR(-E2BIG);
>
> + if (WARN_ON(dest_obj->pages_pin_count == 0))
> + return ERR_PTR(-ENODEV);
> +
> ret = i915_gem_obj_prepare_shmem_read(src_obj, &needs_clflush);
> if (ret) {
> DRM_DEBUG_DRIVER("CMD: failed to prepare shadow batch\n");
> @@ -879,13 +882,6 @@ static u32 *copy_batch(struct drm_i915_gem_object *dest_obj,
> goto unpin_src;
> }
>
> - ret = i915_gem_object_get_pages(dest_obj);
> - if (ret) {
> - DRM_DEBUG_DRIVER("CMD: Failed to get pages for shadow batch\n");
> - goto unmap_src;
> - }
> - i915_gem_object_pin_pages(dest_obj);
> -
> ret = i915_gem_object_set_to_cpu_domain(dest_obj, true);
> if (ret) {
> DRM_DEBUG_DRIVER("CMD: Failed to set shadow batch to CPU\n");
> @@ -895,7 +891,6 @@ static u32 *copy_batch(struct drm_i915_gem_object *dest_obj,
> dst = vmap_batch(dest_obj, 0, batch_len);
> if (!dst) {
> DRM_DEBUG_DRIVER("CMD: Failed to vmap shadow batch\n");
> - i915_gem_object_unpin_pages(dest_obj);
> ret = -ENOMEM;
> goto unmap_src;
> }
> @@ -1126,7 +1121,6 @@ int i915_parse_cmds(struct intel_engine_cs *ring,
> }
>
> vunmap(batch_base);
> - i915_gem_object_unpin_pages(shadow_batch_obj);
>
> return ret;
> }
> diff --git a/drivers/gpu/drm/i915/i915_gem_batch_pool.c b/drivers/gpu/drm/i915/i915_gem_batch_pool.c
> index c690170a1c4f..e5c88ddd8452 100644
> --- a/drivers/gpu/drm/i915/i915_gem_batch_pool.c
> +++ b/drivers/gpu/drm/i915/i915_gem_batch_pool.c
> @@ -66,9 +66,7 @@ void i915_gem_batch_pool_fini(struct i915_gem_batch_pool *pool)
> struct drm_i915_gem_object,
> batch_pool_list);
>
> - WARN_ON(obj->active);
> -
> - list_del_init(&obj->batch_pool_list);
> + list_del(&obj->batch_pool_list);
> drm_gem_object_unreference(&obj->base);
> }
> }
> @@ -96,10 +94,9 @@ i915_gem_batch_pool_get(struct i915_gem_batch_pool *pool,
> WARN_ON(!mutex_is_locked(&pool->dev->struct_mutex));
>
> list_for_each_entry_safe(tmp, next,
> - &pool->cache_list, batch_pool_list) {
> -
> + &pool->cache_list, batch_pool_list) {
> if (tmp->active)
> - continue;
> + break;
>
> /* While we're looping, do some clean up */
> if (tmp->madv == __I915_MADV_PURGED) {
> @@ -113,25 +110,27 @@ i915_gem_batch_pool_get(struct i915_gem_batch_pool *pool,
> * but not 'too much' bigger. A better way to do this
> * might be to bucket the pool objects based on size.
> */
> - if (tmp->base.size >= size &&
> - tmp->base.size <= (2 * size)) {
> + if (tmp->base.size >= size && tmp->base.size <= 2 * size) {
> obj = tmp;
> break;
> }
> }
>
> - if (!obj) {
> + if (obj == NULL) {
> + int ret;
> +
> obj = i915_gem_alloc_object(pool->dev, size);
> - if (!obj)
> + if (obj == NULL)
> return ERR_PTR(-ENOMEM);
>
> - list_add_tail(&obj->batch_pool_list, &pool->cache_list);
> - }
> - else
> - /* Keep list in LRU order */
> - list_move_tail(&obj->batch_pool_list, &pool->cache_list);
> + ret = i915_gem_object_get_pages(obj);
> + if (ret)
> + return ERR_PTR(ret);
>
> - obj->madv = I915_MADV_WILLNEED;
> + obj->madv = I915_MADV_DONTNEED;
> + }
>
> + list_move_tail(&obj->batch_pool_list, &pool->cache_list);
Why is it now safe to do a move_tail instead of add_tail if the node has
just been allocated? Was the original add_tail() wrong or am I not
spotting some critical difference to how new pool objects are created?
> + i915_gem_object_pin_pages(obj);
Is it worth updating the function description comment to add a line
about the returned buffer now being pinned and the caller must worry
about unpinning it?
> return obj;
> }
> diff --git a/drivers/gpu/drm/i915/i915_gem_execbuffer.c b/drivers/gpu/drm/i915/i915_gem_execbuffer.c
> index 3fbd5212225f..6357f01e6c46 100644
> --- a/drivers/gpu/drm/i915/i915_gem_execbuffer.c
> +++ b/drivers/gpu/drm/i915/i915_gem_execbuffer.c
> @@ -37,7 +37,6 @@
> #define __EXEC_OBJECT_HAS_FENCE (1<<30)
> #define __EXEC_OBJECT_NEEDS_MAP (1<<29)
> #define __EXEC_OBJECT_NEEDS_BIAS (1<<28)
> -#define __EXEC_OBJECT_PURGEABLE (1<<27)
>
> #define BATCH_OFFSET_BIAS (256*1024)
>
> @@ -224,12 +223,7 @@ i915_gem_execbuffer_unreserve_vma(struct i915_vma *vma)
> if (entry->flags & __EXEC_OBJECT_HAS_PIN)
> vma->pin_count--;
>
> - if (entry->flags & __EXEC_OBJECT_PURGEABLE)
> - obj->madv = I915_MADV_DONTNEED;
> -
> - entry->flags &= ~(__EXEC_OBJECT_HAS_FENCE |
> - __EXEC_OBJECT_HAS_PIN |
> - __EXEC_OBJECT_PURGEABLE);
> + entry->flags &= ~(__EXEC_OBJECT_HAS_FENCE | __EXEC_OBJECT_HAS_PIN);
> }
>
> static void eb_destroy(struct eb_vmas *eb)
> @@ -1154,6 +1148,7 @@ i915_gem_execbuffer_parse(struct intel_engine_cs *ring,
> batch_start_offset,
> batch_len,
> is_master);
> + i915_gem_object_unpin_pages(shadow_batch_obj);
> if (ret)
> goto err;
>
> @@ -1165,7 +1160,7 @@ i915_gem_execbuffer_parse(struct intel_engine_cs *ring,
>
> vma = i915_gem_obj_to_ggtt(shadow_batch_obj);
> vma->exec_entry = shadow_exec_entry;
> - vma->exec_entry->flags = __EXEC_OBJECT_PURGEABLE | __EXEC_OBJECT_HAS_PIN;
> + vma->exec_entry->flags = __EXEC_OBJECT_HAS_PIN;
> drm_gem_object_reference(&shadow_batch_obj->base);
> list_add_tail(&vma->exec_list, &eb->vmas);
>
More information about the Intel-gfx
mailing list