[Intel-gfx] [CI 02/23] drm/i915: Extract i915_gem_obj_prepare_shmem_write()

Chris Wilson chris at chris-wilson.co.uk
Wed Aug 17 13:55:32 UTC 2016


This is a companion to i915_gem_obj_prepare_shmem_read() that prepares
the backing storage for direct writes. It first serialises with the GPU,
pins the backing storage and then indicates what clfushes are required in
order for the writes to be coherent.

Whilst here, fix support for ancient CPUs without clflush for which we
cannot do the GTT+clflush tricks.

v2: Add i915_gem_obj_finish_shmem_access() for symmetry

Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen at linux.intel.com>
---
 drivers/gpu/drm/i915/i915_cmd_parser.c |   4 +-
 drivers/gpu/drm/i915/i915_drv.h        |  17 +++-
 drivers/gpu/drm/i915/i915_gem.c        | 146 +++++++++++++++++++--------------
 3 files changed, 102 insertions(+), 65 deletions(-)

diff --git a/drivers/gpu/drm/i915/i915_cmd_parser.c b/drivers/gpu/drm/i915/i915_cmd_parser.c
index 1db829c8b912..e586e15e172f 100644
--- a/drivers/gpu/drm/i915/i915_cmd_parser.c
+++ b/drivers/gpu/drm/i915/i915_cmd_parser.c
@@ -973,7 +973,7 @@ static u32 *copy_batch(struct drm_i915_gem_object *dest_obj,
 		       u32 batch_start_offset,
 		       u32 batch_len)
 {
-	int needs_clflush = 0;
+	unsigned int needs_clflush;
 	void *src_base, *src;
 	void *dst = NULL;
 	int ret;
@@ -1020,7 +1020,7 @@ static u32 *copy_batch(struct drm_i915_gem_object *dest_obj,
 unmap_src:
 	vunmap(src_base);
 unpin_src:
-	i915_gem_object_unpin_pages(src_obj);
+	i915_gem_obj_finish_shmem_access(src_obj);
 
 	return ret ? ERR_PTR(ret) : dst;
 }
diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
index 35caa9b2f36a..665039c4c293 100644
--- a/drivers/gpu/drm/i915/i915_drv.h
+++ b/drivers/gpu/drm/i915/i915_drv.h
@@ -3098,9 +3098,6 @@ int i915_gem_object_put_pages(struct drm_i915_gem_object *obj);
 void i915_gem_release_all_mmaps(struct drm_i915_private *dev_priv);
 void i915_gem_release_mmap(struct drm_i915_gem_object *obj);
 
-int i915_gem_obj_prepare_shmem_read(struct drm_i915_gem_object *obj,
-				    int *needs_clflush);
-
 int __must_check i915_gem_object_get_pages(struct drm_i915_gem_object *obj);
 
 static inline int __sg_page_count(struct scatterlist *sg)
@@ -3201,6 +3198,20 @@ static inline void i915_gem_object_unpin_map(struct drm_i915_gem_object *obj)
 	i915_gem_object_unpin_pages(obj);
 }
 
+int i915_gem_obj_prepare_shmem_read(struct drm_i915_gem_object *obj,
+				    unsigned int *needs_clflush);
+int i915_gem_obj_prepare_shmem_write(struct drm_i915_gem_object *obj,
+				     unsigned int *needs_clflush);
+#define CLFLUSH_BEFORE 0x1
+#define CLFLUSH_AFTER 0x2
+#define CLFLUSH_FLAGS (CLFLUSH_BEFORE | CLFLUSH_AFTER)
+
+static inline void
+i915_gem_obj_finish_shmem_access(struct drm_i915_gem_object *obj)
+{
+	i915_gem_object_unpin_pages(obj);
+}
+
 int __must_check i915_mutex_lock_interruptible(struct drm_device *dev);
 int i915_gem_object_sync(struct drm_i915_gem_object *obj,
 			 struct drm_i915_gem_request *to);
diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index a8d0f70c22f9..978ad6e523df 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -609,35 +609,95 @@ __copy_from_user_swizzled(char *gpu_vaddr, int gpu_offset,
  * flush the object from the CPU cache.
  */
 int i915_gem_obj_prepare_shmem_read(struct drm_i915_gem_object *obj,
-				    int *needs_clflush)
+				    unsigned int *needs_clflush)
 {
 	int ret;
 
 	*needs_clflush = 0;
 
-	if (WARN_ON(!i915_gem_object_has_struct_page(obj)))
-		return -EINVAL;
+	if (!i915_gem_object_has_struct_page(obj))
+		return -ENODEV;
 
 	ret = i915_gem_object_wait_rendering(obj, true);
 	if (ret)
 		return ret;
 
-	if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU)) {
-		/* If we're not in the cpu read domain, set ourself into the gtt
-		 * read domain and manually flush cachelines (if required). This
-		 * optimizes for the case when the gpu will dirty the data
-		 * anyway again before the next pread happens. */
+	/* If we're not in the cpu read domain, set ourself into the gtt
+	 * read domain and manually flush cachelines (if required). This
+	 * optimizes for the case when the gpu will dirty the data
+	 * anyway again before the next pread happens.
+	 */
+	if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU))
 		*needs_clflush = !cpu_cache_is_coherent(obj->base.dev,
 							obj->cache_level);
+
+	ret = i915_gem_object_get_pages(obj);
+	if (ret)
+		return ret;
+
+	i915_gem_object_pin_pages(obj);
+
+	if (*needs_clflush && !static_cpu_has(X86_FEATURE_CLFLUSH)) {
+		ret = i915_gem_object_set_to_cpu_domain(obj, false);
+		if (ret) {
+			i915_gem_object_unpin_pages(obj);
+			return ret;
+		}
+		*needs_clflush = 0;
 	}
 
+	return 0;
+}
+
+int i915_gem_obj_prepare_shmem_write(struct drm_i915_gem_object *obj,
+				     unsigned int *needs_clflush)
+{
+	int ret;
+
+	*needs_clflush = 0;
+	if (!i915_gem_object_has_struct_page(obj))
+		return -ENODEV;
+
+	ret = i915_gem_object_wait_rendering(obj, false);
+	if (ret)
+		return ret;
+
+	/* If we're not in the cpu write domain, set ourself into the
+	 * gtt write domain and manually flush cachelines (as required).
+	 * This optimizes for the case when the gpu will use the data
+	 * right away and we therefore have to clflush anyway.
+	 */
+	if (obj->base.write_domain != I915_GEM_DOMAIN_CPU)
+		*needs_clflush |= cpu_write_needs_clflush(obj) << 1;
+
+	/* Same trick applies to invalidate partially written cachelines read
+	 * before writing.
+	 */
+	if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU))
+		*needs_clflush |= !cpu_cache_is_coherent(obj->base.dev,
+							 obj->cache_level);
+
 	ret = i915_gem_object_get_pages(obj);
 	if (ret)
 		return ret;
 
 	i915_gem_object_pin_pages(obj);
 
-	return ret;
+	if (*needs_clflush && !static_cpu_has(X86_FEATURE_CLFLUSH)) {
+		ret = i915_gem_object_set_to_cpu_domain(obj, true);
+		if (ret) {
+			i915_gem_object_unpin_pages(obj);
+			return ret;
+		}
+		*needs_clflush = 0;
+	}
+
+	if ((*needs_clflush & CLFLUSH_AFTER) == 0)
+		obj->cache_dirty = true;
+
+	intel_fb_obj_invalidate(obj, ORIGIN_CPU);
+	obj->dirty = 1;
+	return 0;
 }
 
 /* Per-page copy function for the shmem pread fastpath.
@@ -869,19 +929,14 @@ i915_gem_shmem_pread(struct drm_device *dev,
 	int needs_clflush = 0;
 	struct sg_page_iter sg_iter;
 
-	if (!i915_gem_object_has_struct_page(obj))
-		return -ENODEV;
-
-	user_data = u64_to_user_ptr(args->data_ptr);
-	remain = args->size;
-
-	obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
-
 	ret = i915_gem_obj_prepare_shmem_read(obj, &needs_clflush);
 	if (ret)
 		return ret;
 
+	obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
+	user_data = u64_to_user_ptr(args->data_ptr);
 	offset = args->offset;
+	remain = args->size;
 
 	for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents,
 			 offset >> PAGE_SHIFT) {
@@ -937,7 +992,7 @@ next_page:
 	}
 
 out:
-	i915_gem_object_unpin_pages(obj);
+	i915_gem_obj_finish_shmem_access(obj);
 
 	return ret;
 }
@@ -1242,42 +1297,17 @@ i915_gem_shmem_pwrite(struct drm_device *dev,
 	int shmem_page_offset, page_length, ret = 0;
 	int obj_do_bit17_swizzling, page_do_bit17_swizzling;
 	int hit_slowpath = 0;
-	int needs_clflush_after = 0;
-	int needs_clflush_before = 0;
+	unsigned int needs_clflush;
 	struct sg_page_iter sg_iter;
 
-	user_data = u64_to_user_ptr(args->data_ptr);
-	remain = args->size;
-
-	obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
-
-	ret = i915_gem_object_wait_rendering(obj, false);
+	ret = i915_gem_obj_prepare_shmem_write(obj, &needs_clflush);
 	if (ret)
 		return ret;
 
-	if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
-		/* If we're not in the cpu write domain, set ourself into the gtt
-		 * write domain and manually flush cachelines (if required). This
-		 * optimizes for the case when the gpu will use the data
-		 * right away and we therefore have to clflush anyway. */
-		needs_clflush_after = cpu_write_needs_clflush(obj);
-	}
-	/* Same trick applies to invalidate partially written cachelines read
-	 * before writing. */
-	if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0)
-		needs_clflush_before =
-			!cpu_cache_is_coherent(dev, obj->cache_level);
-
-	ret = i915_gem_object_get_pages(obj);
-	if (ret)
-		return ret;
-
-	intel_fb_obj_invalidate(obj, ORIGIN_CPU);
-
-	i915_gem_object_pin_pages(obj);
-
+	obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
+	user_data = u64_to_user_ptr(args->data_ptr);
 	offset = args->offset;
-	obj->dirty = 1;
+	remain = args->size;
 
 	for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents,
 			 offset >> PAGE_SHIFT) {
@@ -1301,7 +1331,7 @@ i915_gem_shmem_pwrite(struct drm_device *dev,
 		/* If we don't overwrite a cacheline completely we need to be
 		 * careful to have up-to-date data by first clflushing. Don't
 		 * overcomplicate things and flush the entire patch. */
-		partial_cacheline_write = needs_clflush_before &&
+		partial_cacheline_write = needs_clflush & CLFLUSH_BEFORE &&
 			((shmem_page_offset | page_length)
 				& (boot_cpu_data.x86_clflush_size - 1));
 
@@ -1311,7 +1341,7 @@ i915_gem_shmem_pwrite(struct drm_device *dev,
 		ret = shmem_pwrite_fast(page, shmem_page_offset, page_length,
 					user_data, page_do_bit17_swizzling,
 					partial_cacheline_write,
-					needs_clflush_after);
+					needs_clflush & CLFLUSH_AFTER);
 		if (ret == 0)
 			goto next_page;
 
@@ -1320,7 +1350,7 @@ i915_gem_shmem_pwrite(struct drm_device *dev,
 		ret = shmem_pwrite_slow(page, shmem_page_offset, page_length,
 					user_data, page_do_bit17_swizzling,
 					partial_cacheline_write,
-					needs_clflush_after);
+					needs_clflush & CLFLUSH_AFTER);
 
 		mutex_lock(&dev->struct_mutex);
 
@@ -1334,7 +1364,7 @@ next_page:
 	}
 
 out:
-	i915_gem_object_unpin_pages(obj);
+	i915_gem_obj_finish_shmem_access(obj);
 
 	if (hit_slowpath) {
 		/*
@@ -1342,17 +1372,15 @@ out:
 		 * cachelines in-line while writing and the object moved
 		 * out of the cpu write domain while we've dropped the lock.
 		 */
-		if (!needs_clflush_after &&
+		if (!(needs_clflush & CLFLUSH_AFTER) &&
 		    obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
 			if (i915_gem_clflush_object(obj, obj->pin_display))
-				needs_clflush_after = true;
+				needs_clflush |= CLFLUSH_AFTER;
 		}
 	}
 
-	if (needs_clflush_after)
+	if (needs_clflush & CLFLUSH_AFTER)
 		i915_gem_chipset_flush(to_i915(dev));
-	else
-		obj->cache_dirty = true;
 
 	intel_fb_obj_flush(obj, false, ORIGIN_CPU);
 	return ret;
@@ -1431,10 +1459,8 @@ i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
 	if (ret == -EFAULT || ret == -ENOSPC) {
 		if (obj->phys_handle)
 			ret = i915_gem_phys_pwrite(obj, args, file);
-		else if (i915_gem_object_has_struct_page(obj))
-			ret = i915_gem_shmem_pwrite(dev, obj, args, file);
 		else
-			ret = -ENODEV;
+			ret = i915_gem_shmem_pwrite(dev, obj, args, file);
 	}
 
 	i915_gem_object_put(obj);
-- 
2.8.1



More information about the Intel-gfx mailing list